ClawMem — On-device memory layer for Claude Code, OpenClaw, and Hermes agents

On-device memory for Claude Code, OpenClaw, Hermes, and AI agents. Retrieval-augmented search, hooks, and an MCP server in a single local system. No API keys, no cloud dependencies.

ClawMem fuses recent research into a retrieval-augmented memory layer that agents actually use. The hybrid architecture combines QMD-derived multi-signal retrieval (BM25 + vector search + reciprocal rank fusion + query expansion + cross-encoder reranking), SAME-inspired composite scoring (recency decay, confidence, content-type half-lives, co-activation reinforcement), MAGMA-style intent classification with multi-graph traversal (semantic, temporal, and causal beam search), and A-MEM self-evolving memory notes that enrich documents with keywords, tags, and causal links between entries. Pattern extraction from Engram adds deduplication windows, frequency-based durability scoring, and temporal navigation.

Integrates via Claude Code hooks, an MCP server (works with any MCP-compatible client), a native OpenClaw plugin, or a Hermes Agent MemoryProvider plugin. All paths write to the same local SQLite vault. A decision captured during a Claude Code session shows up immediately when an OpenClaw or Hermes agent picks up the same project.

TypeScript on Bun. MIT License.

What It Does

ClawMem turns your markdown notes, project docs, and research dumps into persistent memory for AI coding agents. It automatically:

• Surfaces relevant context on every prompt (context-surfacing hook)
• Bootstraps sessions with your profile, latest handoff, recent decisions, and stale notes
• Captures decisions, preferences, milestones, and problems from session transcripts using a local GGUF observer model
• Imports conversation exports from Claude Code, ChatGPT, Claude.ai, Slack, and plain text via clawmem mine, with optional post-import LLM fact extraction (--synthesize) that pulls structured decisions / preferences / milestones / problems and cross-fact links out of otherwise full-text conversation dumps (v0.7.2)
• Generates handoffs at session end so the next session can pick up where you left off
• Learns what matters via a feedback loop that boosts referenced notes and decays unused ones
• Guards against prompt injection in surfaced content
• Classifies query intent (WHY / WHEN / ENTITY / WHAT) to weight search strategies
• Traverses multi-graphs (semantic, temporal, causal) via adaptive beam search
• Evolves memory metadata as new documents create or refine connections
• Infers causal relationships between facts extracted from session observations
• Detects contradictions between new and prior decisions, auto-decaying superseded ones (with an additional merge-time contradiction gate in the consolidation worker that blocks cross-observation contradictions before they land, v0.7.1)
• Guards against cross-entity merges during consolidation — name-aware dual-threshold merge safety compares entity anchors before merging similar observations, preventing "Alice decided X" from merging into "Bob decided X" (v0.7.1)
• Prevents context bleed in derived insights — the Phase 3 deductive synthesis pipeline validates every draft against an anti-contamination wrapper (deterministic entity contamination check + LLM validator + dedupe) before writing cross-session deductive observations (v0.7.1)
• Frames surfaced facts as background knowledge — context-surfacing wraps injected content in <instruction> + <facts> + <relationships> blocks, telling the model to treat facts as already-known and exposing memory-graph edges between surfaced docs directly in-prompt (v0.7.1)
• Injects knowledge-graph facts as structured triples — when the user's prompt mentions entities already known to the vault, context-surfacing resolves them via a three-path prompt-only extractor (canonical IDs, proper nouns, lowercased n-grams), queries the SPO graph for current-state triples, and appends a <vault-facts> block of raw subject predicate object lines to <vault-context> — off for speed, 200 tokens on balanced, 250 on deep, token-truncated at the triple boundary (v0.9.0)
• Session-scoped focus topic boost — clawmem focus set "<topic>" --session-id <id> writes a per-session focus file that steers query expansion, reranking, chunk selection, snippet extraction, and post-composite-score topic boosting (1.4× match / 0.75× demote) for that session only — session-isolated, fail-open, never writes to SQLite or lifecycle columns (v0.9.0)
• Scores document quality using structure, keywords, and metadata richness signals
• Boosts co-accessed documents — notes frequently surfaced together get retrieval reinforcement
• Decomposes complex queries into typed retrieval clauses (BM25/vector/graph) for multi-topic questions
• Cleans stale embeddings automatically before embed runs, removing orphans from deleted/changed documents
• Transaction-safe indexing — crash mid-index leaves zero partial state (atomic commit with rollback)
• Deduplicates hook-generated observations within a 30-minute window using normalized content hashing, preventing memory bloat from repeated hook output
• Navigates temporal neighborhoods around any document via the timeline tool — progressive disclosure from search to chronological context to full content
• Boosts frequently-revised memories — documents with higher revision counts get a durability signal in composite scoring (capped at 10%)
• Supports pin/snooze lifecycle for persistent boosts and temporary suppression
• Manages document lifecycle — policy-driven archival sweeps with restore capability
• Auto-routes queries via memory_retrieve — classifies intent and dispatches to the optimal search backend
• Syncs project issues from Beads issue trackers into searchable memory
• Runs a quiet-window heavy maintenance lane — a second consolidation worker, off by default behind CLAWMEM_HEAVY_LANE=true, that runs on a longer interval only inside a configurable hour window. Gated by context_usage query-rate so it never competes for CPU/GPU with interactive sessions, scoped exclusively via DB-backed worker_leases, stale-first by default with an optional surprisal selector, and journals every attempt in maintenance_runs for operator visibility (v0.8.0)

Runs fully local with no API keys and no cloud services. Integrates via Claude Code hooks and MCP tools, as a native OpenClaw plugin, or as a Hermes Agent MemoryProvider plugin. All modes share the same vault for cross-runtime memory. Works with any MCP-compatible client.

Full version history is in RELEASE_NOTES.md. Upgrade instructions for existing vaults are in docs/guides/upgrading.md.

Architecture

Install

Platform Support

Platform	Status	Notes
Linux	Full support	Primary target. systemd services for watcher + embed timer.
macOS	Full support	Homebrew SQLite handled automatically. GPU via Metal (llama.cpp).
Windows (WSL2)	Full support	Recommended for Windows users. Install Bun + ClawMem inside WSL2.
Windows (native)	Not recommended	Bun and sqlite-vec work, but bin/clawmem wrapper is bash, hooks expect bash commands, and systemd services have no equivalent. Use WSL2 instead.

Prerequisites

Required:

• Bun v1.0+ — runtime for ClawMem. On Linux, install via curl -fsSL https://bun.sh/install | bash (not snap — snap Bun cannot read stdin, which breaks hooks).
• SQLite with FTS5 — included with Bun. On macOS, install brew install sqlite for extension loading support (ClawMem detects and uses Homebrew SQLite automatically).

Optional (for better performance):

• llama.cpp (llama-server) — for dedicated GPU inference. Without it, node-llama-cpp runs models in-process (auto-downloads on first use). GPU servers give better throughput and prevent silent CPU fallback.
• systemd (Linux) or launchd (macOS) — for persistent background services (watcher, embed timer, GPU servers). ClawMem ships systemd unit templates; macOS users can create equivalent launchd plists. See systemd services.

Optional integrations:

• Claude Code — for hooks + MCP integration
• OpenClaw — for native plugin integration
• Hermes Agent — for MemoryProvider plugin integration
• bd CLI v0.58.0+ — for Beads issue tracker sync (only if using Beads)

Install from npm (recommended)

npm install -g clawmem

If you use Bun as your package manager:

bun add -g clawmem

Install from source

git clone https://github.com/yoloshii/clawmem.git ~/clawmem
cd ~/clawmem && bun install
ln -sf ~/clawmem/bin/clawmem ~/.bun/bin/clawmem

Setup roadmap

After installing, here's the full journey from zero to working memory:

Step	What	How	Details
1. Bootstrap	Create a vault, index your first collection, embed, install hooks and MCP	clawmem bootstrap ~/notes --name notes	One command does it all. Or run each step manually (see below).
2. Choose models	Pick embedding + reranker models based on your hardware	12GB+ VRAM → SOTA stack (zembed-1 + zerank-2). Less → QMD native combo. No GPU → cloud embedding or CPU fallback.	GPU Services
3. Download models	Get the GGUF files for your chosen stack	wget from HuggingFace, or let node-llama-cpp auto-download the QMD native models on first use	Embedding, LLM Server, Reranker Server
4. Start services	Run GPU servers (if using dedicated GPU) and background services. Optionally enable the v0.8.2 background maintenance workers in the watcher unit so consolidation + deductive synthesis run automatically.	llama-server for each model. systemd units for watcher + embed timer. Drop-in for the watcher to enable workers + tune intervals + set the quiet window.	systemd services, background workers
5. Decide what to index	Add collections for your projects, notes, research, and domain docs	clawmem collection add ~/project --name project	The more relevant markdown you index, the better retrieval works. See building a rich context field.
6. Connect your agent	Hook into Claude Code, OpenClaw, Hermes, or any MCP client	clawmem setup hooks && clawmem setup mcp for Claude Code. clawmem setup openclaw for OpenClaw. Copy src/hermes/ to Hermes plugins for Hermes.	Integration
7. Verify	Confirm everything is working	clawmem doctor (full health check) or clawmem status (quick index stats)	Verify Installation

Fastest path: Step 1 alone gets you a working system with in-process CPU/GPU inference and default models — no manual model downloads or service configuration needed. Steps 2-4 are optional upgrades for better performance. Steps 5-6 are where you customize what gets indexed and how your agent connects.

Customize what gets indexed: Each collection has a pattern field in ~/.config/clawmem/config.yaml (default: **/*.md). Tailor it per collection — index project docs, research notes, decision records, Obsidian vaults, or anything else your agents should know about. The more relevant content in the vault, the better retrieval works. See the quickstart for config examples.

Quick start commands

# One command: init + index + embed + hooks + MCP
clawmem bootstrap ~/notes --name notes

# Or step by step:
clawmem init
clawmem collection add ~/notes --name notes
clawmem update --embed
clawmem setup hooks
clawmem setup mcp

# Add more collections (the more you index, the richer retrieval gets)
clawmem collection add ~/projects/myapp --name myapp
clawmem collection add ~/research --name research
clawmem update --embed

# Verify
clawmem doctor

Upgrading

bun update -g clawmem   # or: npm update -g clawmem

Database schema migrates automatically on next startup (new tables and columns are added via CREATE IF NOT EXISTS / ALTER TABLE ADD COLUMN).

After major version updates (e.g. 0.1.x → 0.2.0) that add new enrichment pipelines, run a full enrichment pass to backfill existing documents:

clawmem reindex --enrich  # Full enrichment: entity extraction + links + evolution for all docs
clawmem embed             # Re-embed if upgrading embedding models (not needed for most updates)

--enrich forces the complete A-MEM pipeline (entity extraction, link generation, memory evolution) on all documents, not just new ones. Without it, reindex only refreshes metadata for existing docs.

Routine patch updates (e.g. 0.2.0 → 0.2.1) do not require reindexing.

For version-specific upgrade notes (opt-in features, optional cleanup steps, verification commands), see docs/guides/upgrading.md.

Integration

Claude Code

ClawMem integrates via hooks (settings.json) and an MCP stdio server. Hooks handle 90% of retrieval automatically - the agent never needs to call tools for routine context.

clawmem setup hooks    # Install lifecycle hooks (SessionStart, UserPromptSubmit, Stop, PreCompact)
clawmem setup mcp      # Register MCP server in ~/.claude.json (31 tools)

Automatic (90%): context-surfacing injects relevant memory on every prompt. postcompact-inject re-injects state after compaction. decision-extractor, handoff-generator, feedback-loop capture session state on stop.

Agent-initiated (10%): MCP tools (query, intent_search, find_causal_links, timeline, etc.) for targeted retrieval when hooks don't surface what's needed.

OpenClaw

ClawMem registers as a native OpenClaw memory plugin (kind: memory, v0.10.0+). Same 90/10 automatic retrieval, delivered through OpenClaw's plugin-hook bus instead of Claude Code hooks.

clawmem setup openclaw   # Installs plugin into ~/.openclaw/extensions/clawmem (copy, not symlink)

What the plugin provides:

• before_prompt_build hook (load-bearing) - prompt-aware retrieval (context-surfacing + session-bootstrap) AND the pre-emptive precompact-extract run when token usage approaches the compaction threshold. This is the authoritative path for precompact state capture because it runs synchronously before the LLM call that would trigger compaction, so it cannot race the compactor.
• agent_end hook - decision extraction, handoff generation, feedback loop (parallel, fire-and-forget at the OpenClaw call site). OpenClaw v2026.4.26+ also enforces a 30s default void-hook timeout on agent_end — slow handlers are logged but the underlying postrun work is not cancelled (fail-open).
• before_compaction hook (defense-in-depth fallback) - fires precompact-extract again for the rare case where before_prompt_build's proximity heuristic missed a sudden token-count jump. Fire-and-forget at OpenClaw's call site, so it races the compactor and offers no correctness guarantee on its own — the before_prompt_build path is what actually holds the invariant.
• session_start hook - session registration + cached first-turn bootstrap context
• 5 agent tools - clawmem_search, clawmem_get, clawmem_session_log, clawmem_timeline, clawmem_similar

Disable OpenClaw's native memory search to avoid duplicate injection:

openclaw config set agents.defaults.memorySearch.extraPaths "[]"

ClawMem coexists cleanly with OpenClaw's Active Memory plugin (v2026.4.10+) and, on OpenClaw v2026.4.18+ (#65411), with the memory-core dreaming sidecar — both run alongside ClawMem instead of being mutually exclusive. They search different backends and inject into different prompt regions, so they do not conflict. See the OpenClaw plugin guide — Active Memory coexistence and the memory-core dreaming sidecar section for the two patterns.

Pair ClawMem (memory) with a context-engine plugin (v0.10.0+). OpenClaw and Hermes maintainers have converged on a two-surface plugin model: one slot for memory plugins (cross-session, retrieval-first) and a separate slot for context-engine plugins (in-session, compression/compaction-first). Under that model ClawMem is a memory layer — it has always been one in Hermes via the MemoryProvider ABC, and v0.10.0 moves the OpenClaw integration to the same semantic slot. You can now run ClawMem in the memory slot alongside an LCM-style compression plugin (for example, lossless-claw) in the context-engine slot. The two plugins do not overlap: one persists across sessions, the other reshapes the live window. See the OpenClaw plugin guide — memory vs context engine for the full rationale.

OpenClaw v2026.4.11+ recommended (required for ClawMem v0.10.0+). v2026.4.11 introduced a new plugin discovery contract that requires each plugin directory to ship a package.json with openclaw.extensions declared, and that rejects symlinked plugin directories. ClawMem v0.10.0 includes both fixes. Older ClawMem versions (< v0.10.0) on OpenClaw v2026.4.11+ will fail to discover silently — upgrade ClawMem, then re-run clawmem setup openclaw. See docs/guides/upgrading.md.

Alternative: OpenClaw agents can also use ClawMem's MCP server directly (clawmem setup mcp), with or without hooks. This gives full access to all 31 MCP tools but bypasses OpenClaw's plugin lifecycle, so you lose token budget awareness, native compaction orchestration, and the agent_end message pipeline. The native OpenClaw plugin is recommended for new setups; MCP is available as an additional or standalone integration.

Hermes Agent

ClawMem integrates as a native MemoryProvider plugin — Hermes's pluggable interface for agent memory. Same automatic retrieval and extraction, delivered through Hermes's memory lifecycle instead of Claude Code hooks.

Install:

# Preferred — user-plugin path (Hermes #10529, v2026.4.13+).
# Survives `git pull` of hermes-agent and avoids dual-registration with bundled providers.
cp -r /path/to/ClawMem/src/hermes ${HERMES_HOME:-~/.hermes}/plugins/clawmem

# Or, the bundled-style path (always supported, takes precedence on name collisions).
# Recommended only when you actively work in the hermes-agent source tree.
cp -r /path/to/ClawMem/src/hermes /path/to/hermes-agent/plugins/memory/clawmem

# Symlink alternative for in-place development (either path).
ln -s /path/to/ClawMem/src/hermes ${HERMES_HOME:-~/.hermes}/plugins/clawmem

Configure in your Hermes profile's .env or environment:

CLAWMEM_BIN=/path/to/clawmem          # Path to clawmem binary (or ensure it's on PATH)
CLAWMEM_SERVE_PORT=7438                # REST API port (default: 7438)
CLAWMEM_SERVE_MODE=external            # "external" (you run clawmem serve) or "managed" (plugin manages it)
CLAWMEM_PROFILE=balanced               # speed | balanced | deep

Then set memory.provider: clawmem in your Hermes config.yaml, or run hermes memory setup to configure interactively.

What the plugin provides:

• prefetch() — prompt-aware retrieval via context-surfacing hook (automatic every turn)
• on_session_end() — decision extraction, handoff generation, feedback loop (parallel)
• on_pre_compress() — pre-compaction state preservation
• session-bootstrap — session registration + first-turn context injection
• 5 agent tools — clawmem_retrieve, clawmem_get, clawmem_session_log, clawmem_timeline, clawmem_similar
• Plugin-managed transcript — maintains its own JSONL transcript for ClawMem hooks

Requirements: clawmem binary on PATH and clawmem serve running (external mode) or the plugin starts it automatically (managed mode). Python 3.10+. No pip dependencies beyond Hermes itself (uses urllib for REST calls, httpx optional for better performance).

Alternative: Hermes also has built-in MCP client support. You can add ClawMem as an MCP server in Hermes's config.yaml under mcp_servers for tool-only access. But this misses the lifecycle hooks (prefetch, session_end, pre_compress), so the native plugin is recommended.

See Hermes plugin guide for architecture details, lifecycle mapping, and troubleshooting.

Multi-Framework Operation

All three integrations share the same SQLite vault by default. Claude Code, OpenClaw, and Hermes can run simultaneously — decisions captured in one runtime are immediately available in the others, giving agents persistent shared memory across sessions and platforms. WAL mode + busy_timeout handles concurrent access.

Multi-Vault (Optional)

By default, ClawMem uses a single vault at ~/.cache/clawmem/index.sqlite. For users who want separate memory domains (e.g., work vs personal, or isolated vaults per project), ClawMem supports named vaults.

Configure in ~/.config/clawmem/config.yaml:

vaults:
  work: ~/.cache/clawmem/work.sqlite
  personal: ~/.cache/clawmem/personal.sqlite

Or via environment variable:

export CLAWMEM_VAULTS='{"work":"~/.cache/clawmem/work.sqlite","personal":"~/.cache/clawmem/personal.sqlite"}'

Using vaults with MCP tools:

All retrieval tools (memory_retrieve, query, search, vsearch, intent_search) accept an optional vault parameter. Omit it to use the default vault.

# Search the default vault (no vault param needed)
query("authentication flow")

# Search a named vault
query("project timeline", vault="work")

# List configured vaults
list_vaults()

# Sync content into a vault
vault_sync(vault="work", content_root="~/work/docs")

Single-vault users: No action needed. Everything works without configuration. The vault parameter is always optional and ignored when no vaults are configured.

GPU Services

ClawMem uses three llama-server (llama.cpp) instances for neural inference. All three have in-process fallbacks via node-llama-cpp (auto-downloads on first use), so ClawMem works without a dedicated GPU. node-llama-cpp auto-detects the best available backend — Metal on Apple Silicon, Vulkan where available, CPU as last resort. With GPU acceleration (Metal/Vulkan), in-process inference is fast for these small models (0.3B–1.7B); on CPU-only systems it is significantly slower. For production use, run the servers via systemd services to prevent silent fallback.

GPU with VRAM to spare (12GB+, recommended): ZeroEntropy's distillation-paired stack delivers best retrieval quality — total ~10GB VRAM.

Service	Port	Model	VRAM	Purpose
Embedding	8088	zembed-1-Q4_K_M	~4.4GB	SOTA embedding (2560d, 32K context). Distilled from zerank-2 via zELO.
LLM	8089	qmd-query-expansion-1.7B-q4_k_m	~2.2GB	Intent classification, query expansion, A-MEM
Reranker	8090	zerank-2-Q4_K_M	~3.3GB	SOTA reranker. Outperforms Cohere rerank-3.5. Optimal pairing with zembed-1.

Important: zembed-1 and zerank-2 use non-causal attention — -ub must equal -b on llama-server (e.g. -b 2048 -ub 2048). See Reranker Server for details.

License: zembed-1 and zerank-2 are released under CC-BY-NC-4.0 — non-commercial only. The QMD native models below have no such restriction.

No dedicated GPU / GPU without VRAM to spare: The QMD native combo — total ~4GB VRAM, also runs via node-llama-cpp (Metal on Apple Silicon, Vulkan where available, CPU as last resort). Fast with GPU acceleration; significantly slower on CPU-only.

Service	Port	Model	VRAM	Purpose
Embedding	8088	EmbeddingGemma-300M-Q8_0	~400MB	Vector search, indexing, context-surfacing (768d, 2K context)
LLM	8089	qmd-query-expansion-1.7B-q4_k_m	~2.2GB	Intent classification, query expansion, A-MEM
Reranker	8090	qwen3-reranker-0.6B-Q8_0	~1.3GB	Cross-encoder reranking (query, intent_search)

The bin/clawmem wrapper defaults to localhost:8088/8089/8090. If a server is unreachable (transport error like ECONNREFUSED/ETIMEDOUT), ClawMem sets a 60-second cooldown and falls back to in-process inference via node-llama-cpp (auto-downloads the QMD native models on first use, uses Metal/Vulkan/CPU depending on hardware). HTTP errors (400/500) and user-cancelled requests do not trigger cooldown — the remote server is retried normally on the next call. With GPU acceleration the fallback is fast; on CPU-only it is significantly slower. ClawMem always works either way, but if you're running dedicated GPU servers, use systemd services to ensure they stay up.

To prevent fallback and fail fast instead, set CLAWMEM_NO_LOCAL_MODELS=true.

Remote GPU (optional)

If your GPU lives on a separate machine, point the env vars at it:

export CLAWMEM_EMBED_URL=http://gpu-host:8088
export CLAWMEM_LLM_URL=http://gpu-host:8089
export CLAWMEM_LLM_MODEL=qwen3
export CLAWMEM_RERANK_URL=http://gpu-host:8090

For remote setups, set CLAWMEM_NO_LOCAL_MODELS=true to prevent node-llama-cpp from auto-downloading multi-GB model files if a server is unreachable.

No Dedicated GPU (in-process inference)

All three QMD native models run locally without a dedicated GPU. node-llama-cpp auto-downloads them on first use (~300MB embedding + ~1.1GB LLM + ~600MB reranker) and auto-detects the best backend — Metal on Apple Silicon (fast, uses integrated GPU), Vulkan where available (fast, uses discrete or integrated GPU), or CPU as last resort (significantly slower). With Metal or Vulkan, in-process inference handles these small models well; CPU-only is functional but noticeably slower.

Alternatively, use a cloud embedding provider if you prefer not to run models locally.

Embedding

ClawMem calls the OpenAI-compatible /v1/embeddings endpoint for all embedding operations. This works with local llama-server instances and cloud providers alike.

Option A: GPU with VRAM to spare (recommended)

Use zembed-1-Q4_K_M — SOTA retrieval quality, distilled from zerank-2 via ZeroEntropy's zELO methodology. CC-BY-NC-4.0 — non-commercial only.

• Size: 2.4GB, Dimensions: 2560, VRAM: ~4.4GB, Context: 32K tokens

wget https://huggingface.co/Abhiray/zembed-1-Q4_K_M-GGUF/resolve/main/zembed-1-Q4_K_M.gguf

# -ub must match -b for non-causal attention
llama-server -m zembed-1-Q4_K_M.gguf \
  --embeddings --port 8088 --host 0.0.0.0 \
  -ngl 99 -c 8192 -b 2048 -ub 2048

Option B: No GPU / GPU without VRAM to spare

Use EmbeddingGemma-300M-Q8_0 — the QMD native embedding model. Only 300MB, runs on CPU or any GPU.

• Size: 314MB, Dimensions: 768, VRAM: ~400MB (or CPU), Context: 2048 tokens

wget https://huggingface.co/ggml-org/embeddinggemma-300M-GGUF/resolve/main/embeddinggemma-300M-Q8_0.gguf

# On GPU (add -ngl 99):
llama-server -m embeddinggemma-300M-Q8_0.gguf \
  --embeddings --port 8088 --host 0.0.0.0 \
  -ngl 99 -c 2048 --batch-size 2048

# On CPU (omit -ngl):
llama-server -m embeddinggemma-300M-Q8_0.gguf \
  --embeddings --port 8088 --host 0.0.0.0 \
  -c 2048 --batch-size 2048

For multilingual corpora, the SOTA zembed-1 (Option A) supports multilingual out of the box. For a lightweight alternative: granite-embedding-278m-multilingual-Q6_K (314MB, set CLAWMEM_EMBED_MAX_CHARS=1100 due to 512-token context).

Option C: Cloud Embedding API

Alternatively, use a cloud embedding provider instead of running a local server. Any provider with an OpenAI-compatible /v1/embeddings endpoint works.

Configuration: Copy .env.example to .env and set your provider credentials:

cp .env.example .env
# Edit .env:
CLAWMEM_EMBED_URL=https://api.jina.ai
CLAWMEM_EMBED_API_KEY=jina_your-key-here
CLAWMEM_EMBED_MODEL=jina-embeddings-v5-text-small

Or export them in your shell. Precedence: shell environment > .env file > bin/clawmem wrapper defaults.

Provider	CLAWMEM_EMBED_URL	CLAWMEM_EMBED_MODEL	Dimensions	Notes
Jina AI	https://api.jina.ai	jina-embeddings-v5-text-small	1024	32K context, task-specific LoRA adapters
OpenAI	https://api.openai.com	text-embedding-3-small	1536	8K context, Matryoshka dimensions via CLAWMEM_EMBED_DIMENSIONS
Voyage AI	https://api.voyageai.com	voyage-4-large	1024	32K context
Cohere	https://api.cohere.com	embed-v4.0	1024	128K context

Cloud mode auto-detects your provider from the URL and sends the right parameters (Jina task, Voyage/Cohere input_type, OpenAI dimensions). Batch embedding (50 fragments/request), server-side truncation, adaptive TPM-aware pacing, and retry with jitter are all handled automatically. Set CLAWMEM_EMBED_TPM_LIMIT to match your provider tier (default: 100000). See docs/guides/cloud-embedding.md for full details.

Note: Cloud providers handle their own context window limits — ClawMem skips client-side truncation when an API key is set. Local llama-server truncates at CLAWMEM_EMBED_MAX_CHARS (default: 6000 chars).

Verify and embed

# Verify endpoint is reachable
curl $CLAWMEM_EMBED_URL/v1/embeddings \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $CLAWMEM_EMBED_API_KEY" \
  -d "{\"input\":\"test\",\"model\":\"$CLAWMEM_EMBED_MODEL\"}"

# Embed your vault
./bin/clawmem embed

LLM Server

Intent classification, query expansion, and A-MEM extraction use qmd-query-expansion-1.7B — a Qwen3-1.7B finetuned by QMD specifically for generating search expansion terms (hyde, lexical, and vector variants). ~1.1GB at q4_k_m quantization, served via llama-server on port 8089.

Without a server: If CLAWMEM_LLM_URL is unset, node-llama-cpp auto-downloads the model on first use.

Performance (RTX 3090):

• Intent classification: 27ms
• Query expansion: 333 tok/s
• VRAM: ~2.2-2.8GB depending on quantization

Qwen3 /no_think flag: Qwen3 uses thinking tokens by default. ClawMem appends /no_think to all prompts automatically to get structured output in the content field.

Intent classification: Uses a dual-path approach:

1. Heuristic regex classifier (instant) — handles strong signals (why/when/who keywords) with 0.8+ confidence
2. LLM refinement (27ms on GPU) — only for ambiguous queries below 0.8 confidence

Server setup:

# Download the finetuned model
wget https://huggingface.co/tobil/qmd-query-expansion-1.7B-gguf/resolve/main/qmd-query-expansion-1.7B-q4_k_m.gguf

# Start llama-server for LLM inference
llama-server -m qmd-query-expansion-1.7B-q4_k_m.gguf \
  --port 8089 --host 0.0.0.0 \
  -ngl 99 -c 4096 --batch-size 512

Reranker Server

Cross-encoder reranking for query and intent_search pipelines on port 8090. ClawMem calls the /v1/rerank endpoint (or falls back to scoring via /v1/completions for compatible servers).

Scores each candidate against the original query (cross-encoder architecture). query pipeline: 4000 char context per doc (deep reranking); intent_search: 200 char context per doc (fast reranking).

GPU with VRAM to spare (recommended): zerank-2-Q4_K_M (2.4GB, ~3.3GB VRAM). Outperforms Cohere rerank-3.5 and Gemini 2.5 Flash. Optimal pairing with zembed-1 (same distillation architecture via zELO). CC-BY-NC-4.0 — non-commercial only.

wget https://huggingface.co/keisuke-miyako/zerank-2-gguf-q4_k_m/resolve/main/zerank-2-Q4_k_m.gguf

# -ub must match -b for non-causal attention
llama-server -m zerank-2-Q4_K_M.gguf \
  --reranking --port 8090 --host 0.0.0.0 \
  -ngl 99 -c 2048 -b 2048 -ub 2048

CPU / GPU without VRAM to spare: qwen3-reranker-0.6B-Q8_0 (~600MB, ~1.3GB VRAM). The QMD native reranker — auto-downloaded by node-llama-cpp if no server is running.

wget https://huggingface.co/ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/resolve/main/Qwen3-Reranker-0.6B-Q8_0.gguf

llama-server -m Qwen3-Reranker-0.6B-Q8_0.gguf \
  --reranking --port 8090 --host 0.0.0.0 \
  -ngl 99 -c 2048 --batch-size 512

Note: zerank-2 and zembed-1 use non-causal attention — -ub (ubatch) must equal -b (batch). Omitting -ub or setting it lower causes assertion crashes. qwen3-reranker-0.6B does not have this requirement. See llama.cpp#12836.

MCP Server

ClawMem exposes 31 MCP tools via the Model Context Protocol and an optional HTTP REST API. Any MCP-compatible client or HTTP client can use it.

Claude Code (automatic):

./bin/clawmem setup mcp   # Registers in ~/.claude.json

Manual (any MCP client):

Add to your MCP config (e.g. ~/.claude.json, claude_desktop_config.json, or your client's equivalent):

{
  "mcpServers": {
    "clawmem": {
      "command": "/absolute/path/to/clawmem/bin/clawmem",
      "args": ["mcp"]
    }
  }
}

The server runs via stdio — no network port needed. The bin/clawmem wrapper sets the GPU endpoint env vars automatically.

Verify: After registering, your client should see tools including memory_retrieve, search, vsearch, query, query_plan, intent_search, timeline, etc.

HTTP REST API (optional)

For web dashboards, non-MCP agents, cross-machine access, or programmatic use:

./bin/clawmem serve                          # localhost:7438, no auth
./bin/clawmem serve --port 8080              # custom port
CLAWMEM_API_TOKEN=secret ./bin/clawmem serve # with bearer token auth

Endpoints:

Method	Path	Description
GET	/health	Liveness probe + version + doc count
GET	/stats	Full index statistics
POST	/search	Unified search (mode: auto/keyword/semantic/hybrid)
POST	/retrieve	Smart retrieve with auto-routing (mode: auto/keyword/semantic/causal/timeline/hybrid)
GET	/documents/:docid	Single document by 6-char hash prefix
GET	/documents?pattern=...	Multi-get by glob pattern
GET	/timeline/:docid	Temporal neighborhood (before/after)
GET	/sessions	Recent session history
GET	/collections	List all collections
GET	/lifecycle/status	Active/archived/pinned/snoozed counts
POST	/documents/:docid/pin	Pin/unpin
POST	/documents/:docid/snooze	Snooze until date
POST	/documents/:docid/forget	Deactivate
POST	/lifecycle/sweep	Archive stale docs (dry_run default)
GET	/graph/causal/:docid	Causal chain traversal
GET	/graph/similar/:docid	k-NN neighbors
GET	/export	Full vault export as JSON
POST	/reindex	Trigger re-scan
POST	/graphs/build	Rebuild temporal + semantic graphs

Auth: Set CLAWMEM_API_TOKEN env var to require Authorization: Bearer <token> on all requests. If unset, access is open (localhost-only by default). See .env.example.

Search example:

curl -X POST http://localhost:7438/search \
  -H 'Content-Type: application/json' \
  -d '{"query": "authentication decisions", "mode": "hybrid", "compact": true}'

Verify Installation

./bin/clawmem doctor   # Full health check
./bin/clawmem status   # Quick index status
bun test               # Run test suite

Agent Instructions

ClawMem ships three instruction files and an optional maintenance agent:

File	Loaded	Purpose
CLAUDE.md	Automatically (Claude Code, when working in this repo)	Complete operational reference — hooks, tools, query optimization, scoring, pipeline details, troubleshooting
AGENTS.md	Framework-dependent	Identical to CLAUDE.md — cross-framework compatibility (Cursor, Windsurf, Codex, etc.)
SKILL.md	On-demand (agent reads when needed)	Same reference as CLAUDE.md, shipped with the package for cross-project use
agents/clawmem-curator.md	On-demand via clawmem setup curator	Maintenance agent — lifecycle triage, retrieval health checks, dedup sweeps, graph rebuilds

Working in the ClawMem repo: No action needed — CLAUDE.md loads automatically.

Using ClawMem from other projects: Your agent needs instructions on how to use ClawMem's hooks and MCP tools. Two options:

Option A: Copy instructions into your project

Copy the contents of CLAUDE.md (or the relevant sections) into your project's own CLAUDE.md or AGENTS.md. Simple but requires manual updates when ClawMem changes.

Option B: Add a trigger block (recommended)

Add this minimal trigger block to your global ~/.claude/CLAUDE.md. It gives the agent routing rules always loaded, and tells it how to find the full reference (SKILL.md) shipped with your installation when deeper guidance is needed:

## ClawMem

Architecture: hooks (automatic, ~90%) + MCP tools (explicit, ~10%).

Vault: `~/.cache/clawmem/index.sqlite` | Config: `~/.config/clawmem/config.yaml`

### Escalation Gate (3 rules — ONLY escalate to MCP tools when one fires)

1. **Low-specificity injection** — `<vault-context>` is empty or lacks the specific fact needed
2. **Cross-session question** — "why did we decide X", "what changed since last time"
3. **Pre-irreversible check** — before destructive or hard-to-reverse changes

### Tool Routing (once escalated)

**Preferred:** `memory_retrieve(query)` — auto-classifies and routes to the optimal backend.

**Direct routing** (when calling specific tools):

    "why did we decide X"         → intent_search(query)          NOT query()
    "what happened last session"  → session_log()                 NOT query()
    "what else relates to X"      → find_similar(file)            NOT query()
    Complex multi-topic           → query_plan(query)             NOT query()
    General recall                → query(query, compact=true)
    Keyword spot check            → search(query, compact=true)
    Conceptual/fuzzy              → vsearch(query, compact=true)
    Full content                  → multi_get("path1,path2")
    Lifecycle health              → lifecycle_status()
    Stale sweep                   → lifecycle_sweep(dry_run=true)
    Restore archived              → lifecycle_restore(query)

ALWAYS `compact=true` first → review → `multi_get` for full content.

### Proactive Use (no escalation gate needed)

- User says "remember this" / critical decision made → `memory_pin(query)` immediately
- User corrects a misconception → `memory_pin(query)` the correction
- `<vault-context>` surfaces irrelevant/noisy content → `memory_snooze(query, until)` for 30 days
- Need to correct a memory → `memory_forget(query)`
- After bulk ingestion → `build_graphs`

### Anti-Patterns

- Do NOT use `query()` for everything — match query type to tool, or use `memory_retrieve`
- Do NOT call query/intent_search every turn — 3 rules above are the only gates
- Do NOT re-search what's already in `<vault-context>`
- Do NOT pin everything — pin is for persistent high-priority items, not routine decisions
- Do NOT forget memories to "clean up" — let confidence decay handle it
- Do NOT wait for curator to pin decisions — pin immediately when critical

For detailed operational guidance (query optimization, troubleshooting, collection setup, embedding workflow, graph building, curator), find and read the shipped SKILL.md:
  Bash: CLAWMEM_ROOT=$(cd "$(dirname "$(which clawmem)")/.." && pwd) && echo "$CLAWMEM_ROOT/SKILL.md"
  Then: Read the file at that path.

This gives your agent the 3-rule gate, tool routing, and proactive behaviors always loaded. When it needs deeper guidance, it locates and reads the full SKILL.md reference shipped with your installation — no symlinks or skill registration required.

---

CLI Reference

clawmem init                                    Create DB + config
clawmem bootstrap <vault> [--name N] [--skip-embed]  One-command setup
clawmem collection add <path> --name <name>     Add a collection
clawmem collection list                         List collections
clawmem collection remove <name>                Remove a collection

clawmem update [--pull] [--embed]               Incremental re-scan
clawmem mine <dir> [-c name] [--embed] [--synthesize]  Import conversation exports (--synthesize runs post-import LLM fact extraction, v0.7.2)
clawmem embed [-f]                              Generate fragment embeddings
clawmem reindex [--force]                       Full re-index
clawmem watch                                   File watcher daemon

clawmem search <query> [-n N] [--json]          BM25 keyword search
clawmem vsearch <query> [-n N] [--json]         Vector semantic search
clawmem query <query> [-n N] [--json]           Full hybrid pipeline

clawmem profile                                 Show user profile
clawmem profile rebuild                         Force profile rebuild
clawmem update-context                          Regenerate per-folder CLAUDE.md

clawmem list [-n/--limit N] [-c col] [--json]   Browse recent documents
clawmem budget [--session ID]                   Token utilization
clawmem log [--last N]                          Session history
clawmem hook <name>                             Manual hook trigger
clawmem surface --context --stdin               IO6: pre-prompt context injection
clawmem surface --bootstrap --stdin             IO6: per-session bootstrap injection

clawmem reflect [N]                             Cross-session reflection (last N days, default 14)
clawmem consolidate [--dry-run] [N]             Find and archive duplicate low-confidence docs

clawmem install-service [--enable] [--remove]   Systemd watcher service
clawmem setup hooks [--remove]                  Install/remove Claude Code hooks
clawmem setup mcp [--remove]                    Register/remove MCP server
clawmem setup curator [--remove]                Install/remove curator maintenance agent
clawmem mcp                                     Start stdio MCP server
clawmem serve [--port 7438] [--host 127.0.0.1]  Start HTTP REST API server
clawmem path                                    Print database path
clawmem doctor                                  Full health check
clawmem status                                  Quick index status

MCP Tools (31)

Registered by clawmem setup mcp. Available to any MCP-compatible client.

Tool	Description
__IMPORTANT	Workflow guide: prefer memory_retrieve → match query type to tool → multi_get for full content

Core Search & Retrieval

Tool	Description
memory_retrieve	Preferred entry point. Auto-classifies query and routes to optimal backend (query, intent_search, session_log, find_similar, or query_plan). Use instead of manually choosing a search tool.
search	BM25 keyword search — for exact terms, config names, error codes, filenames. Composite scoring + co-activation boost + compact mode. Collection filter supports comma-separated values. Prefer memory_retrieve for auto-routing.
vsearch	Vector semantic search — for conceptual/fuzzy matching when exact keywords are unknown. Composite scoring + co-activation boost + compact mode. Collection filter supports comma-separated values. Prefer memory_retrieve for auto-routing.
query	Full hybrid pipeline (BM25 + vector + rerank) — general-purpose when query type is unclear. WRONG for "why" questions (use intent_search) or cross-session queries (use session_log). Prefer memory_retrieve for auto-routing. Intent hint, strong-signal bypass, chunk dedup, candidateLimit, MMR diversity, compact mode.
get	Retrieve single document by path or docid
multi_get	Retrieve multiple docs by glob or comma-separated list
find_similar	USE THIS for "what else relates to X", "show me similar docs". Finds k-NN vector neighbors — discovers connections beyond keyword overlap that search/query cannot find.

Intent-Aware Search

Tool	Description
intent_search	USE THIS for "why did we decide X", "what caused Y", "who worked on Z". Classifies intent (WHY/WHEN/ENTITY/WHAT), traverses causal + semantic graph edges. Returns decision chains that query() cannot find.
query_plan	USE THIS for complex multi-topic queries ("tell me about X and also Y", "compare A with B"). Decomposes into parallel typed clauses (bm25/vector/graph), executes each, merges via RRF. query() searches as one blob — this tool splits topics and routes each optimally.

intent_search pipeline: Query → Intent Classification → BM25 + Vector → Intent-Weighted RRF → Graph Expansion (WHY/ENTITY intents) → Cross-Encoder Reranking → Composite Scoring

query_plan pipeline: Query → LLM decomposition into 2-4 typed clauses → Parallel execution (BM25/vector/graph per clause) → RRF merge across clauses → Composite scoring. Falls back to single-query for simple inputs.

Multi-Graph & Causal

Tool	Description
build_graphs	Build temporal and/or semantic graphs from document corpus
find_causal_links	Trace decision chains: "what led to X", "how we got from A to B". Follow up intent_search with this tool on a top result to walk the full causal chain. Traverses causes / caused_by / both up to N hops with depth-annotated reasoning.
kg_query	Query the SPO knowledge graph: "what does X relate to?", "what was true about X when?". Returns temporal entity-relationship triples with validity windows. Accepts entity name (resolved via searchEntities) or canonical ID in vault:type:slug form. Triples are populated by the decision-extractor hook from observer-emitted <triples> blocks.
memory_evolution_status	Show how a document's A-MEM metadata evolved over time
timeline	Show the temporal neighborhood around a document — what was created/modified before and after it. Progressive disclosure: search → timeline (context) → get (full content). Supports same-collection scoping and session correlation.

Beads Integration

Tool	Description
beads_sync	Sync Beads issues from Dolt backend (bd CLI) into memory: creates docs, bridges all dep types to memory_relations, runs A-MEM enrichment

Vault Management

Tool	Description
list_vaults	Show configured vault names and paths. Empty in single-vault mode.
vault_sync	Index markdown from a directory into a named vault. Restricted-path validation rejects sensitive directories.

Agent Diary

Tool	Description
diary_write	Write a diary entry. Use for recording important events, decisions, or observations in environments without hook support. Stored as searchable memories.
diary_read	Read recent diary entries. Filter by agent name.

Memory Management & Lifecycle

Tool	Description
memory_forget	Search → deactivate closest match (with audit trail)
memory_pin	Pin a memory for +0.3 composite boost. USE PROACTIVELY when: user states a persistent constraint, makes an architecture decision, or corrects a misconception. Don't wait for curator — pin critical decisions immediately.
memory_snooze	Temporarily hide a memory from context surfacing until a date. USE PROACTIVELY when <vault-context> repeatedly surfaces irrelevant content — snooze for 30 days instead of ignoring it.
status	Index health with content type distribution
reindex	Trigger vault re-scan
index_stats	Detailed stats: types, staleness, access counts, sessions
session_log	USE THIS for "last time", "yesterday", "what happened", "what did we do". Returns session history with handoffs and file changes. DO NOT use query() for cross-session questions — this tool has session-specific data that search cannot find.
profile	Current static + dynamic user profile
lifecycle_status	Document lifecycle statistics: active, archived, forgotten, pinned, snoozed counts and policy summary
lifecycle_sweep	Run lifecycle policies: archive stale docs past retention threshold, optionally purge old archives. Defaults to dry_run (preview only)
lifecycle_restore	Restore documents that were auto-archived by lifecycle policies. Filter by query, collection, or restore all

Compact Mode

search, vsearch, and query accept compact: true to return { id, path, title, score, snippet, content_type, fragment } instead of full content. Saves ~5x tokens for initial filtering.

Hooks (Claude Code Integration)

Hooks installed by clawmem setup hooks:

Hook	Event	What It Does
context-surfacing	UserPromptSubmit	Hybrid search → FTS supplement → file-aware search (E13) → snooze filter → spreading activation (E11) → memory type diversification (E10) → tiered injection (HOT/WARM/COLD) → <vault-context> + <vault-routing> hint. Profile-driven budget/results/timeout.
postcompact-inject	SessionStart	Re-injects authoritative context after compaction: precompact state + recent decisions + antipatterns + vault context (1200 token budget)
curator-nudge	SessionStart	Surfaces curator report actions, nudges when report is stale (>7 days)
precompact-extract	PreCompact	Extracts decisions, file paths, open questions before auto-compaction → writes precompact-state.md to auto-memory
decision-extractor	Stop	GGUF observer extracts structured observations (decisions, preferences, milestones, problems, bugfixes, features, refactors, discoveries), infers causal links, detects contradictions with prior decisions
handoff-generator	Stop	GGUF observer generates rich handoff, regex fallback
feedback-loop	Stop	Silently boosts referenced notes, decays unused ones, records co-activation + usage relations between co-referenced docs, tracks utility signals (surfaced vs referenced ratio for lifecycle automation)

Additional hooks available but not installed by default:

Hook	Event	Why Not Default
session-bootstrap	SessionStart	Injects ~2000 tokens before user types anything. context-surfacing on first prompt is more precise.
staleness-check	SessionStart	Redundant without session-bootstrap (stale notes are part of its output).
pretool-inject	PreToolUse	Disabled in HOOK_EVENT_MAP (cannot inject additionalContext via PreToolUse).

Hooks handle ~90% of retrieval automatically. For agent escalation logic (when to use MCP tools vs rely on hooks), see CLAUDE.md.

Search Pipeline

User Query + optional intent hint
  → BM25 Probe → Strong Signal Check (skip expansion if top hit ≥ 0.85 with gap ≥ 0.15; disabled when intent provided)
  → Query Expansion (intent steers LLM prompt when provided)
  → BM25 + Vector Search (parallel, original query 2× weight)
  → Reciprocal Rank Fusion → slice to candidateLimit (default 30)
  → Intent-Aware Chunk Selection (intent terms at 0.5× weight alongside query terms at 1.0×)
  → Cross-Encoder Reranking (4000 char context; intent prepended; chunk dedup; batch cap=4)
  → Position-Aware Blending (α=0.75 top3, 0.60 mid, 0.40 tail)
  → SAME Composite Scoring ((search × 0.5 + recency × 0.25 + confidence × 0.25) × qualityMultiplier × lengthNorm × coActivationBoost + pinBoost)
  → MMR Diversity Filter (Jaccard bigram similarity > 0.6 → demoted)
  → Ranked Results

For agent-facing query optimization (tool selection, query string quality, intent parameter, candidateLimit), see CLAUDE.md.

Multi-Graph Traversal

For WHY and ENTITY queries, the search pipeline expands results through the memory graph:

1. Start from top-10 baseline results as anchor nodes
2. For each frontier node: get neighbors via any relation type
3. Score transitions: λ1·structure + λ2·semantic_affinity
4. Apply decay: new_score = parent_score * γ + transition_score
5. Keep top-k (beam search), repeat until max depth or budget

Graph types:

• Semantic — vector similarity edges (threshold > 0.7)
• Temporal — chronological document ordering
• Causal — LLM-inferred cause→effect from Observer facts + Beads blocks/waits-for deps
• Supporting — LLM-analyzed document relationships + Beads discovered-from deps
• Contradicts — LLM-analyzed document relationships. Additional contradicts edges are inserted by the consolidation worker's merge-time contradiction gate (v0.7.1) when it blocks a Phase 2 merge under the link policy, and by Phase 3 deductive synthesis when a new draft contradicts a prior deductive observation.

Content Type Scoring

Type	Half-life	Baseline	Notes
decision	∞	0.85	Never decays
deductive	∞	0.85	Never decays — cross-session derived insights with source provenance
preference	∞	0.80	Never decays — user preferences are durable facts
hub	∞	0.80	Never decays
antipattern	∞	0.75	Never decays — accumulated negative patterns persist
problem	60 days	0.75	High priority but resolves over time
research	90 days	0.70
milestone	60 days	0.70	Important at the time, fades as project moves forward
project	120 days	0.65
handoff	30 days	0.60	Fast decay — most recent matters
conversation	45 days	0.55	Imported chat exchanges
progress	45 days	0.50
note	60 days	0.50	Default

Content types are inferred from frontmatter or file path patterns. Half-lives extend up to 3× for frequently-accessed memories (access reinforcement, decays over 90 days). Non-durable types (handoff, progress, conversation, note, project) lose 5% confidence per week without access (attention decay). Decision/deductive/preference/hub/research/antipattern are exempt.

Quality scoring: Each document gets a quality_score (0.0–1.0) computed during indexing based on length, structure (headings, lists), decision/correction keywords, and frontmatter richness. Applied as qualityMultiplier = 0.7 + 0.6 × qualityScore (range: 0.7× penalty to 1.3× boost).

Length normalization: 1/(1 + 0.5 × log2(max(bodyLength/500, 1))) — penalizes verbose entries that dominate via keyword density. Floor at 30% of original score.

Frequency boost: Documents with higher revision counts or duplicate counts get a durability signal: freqSignal = (revisions - 1) × 2 + (duplicates - 1), freqBoost = min(0.10, log1p(freqSignal) × 0.03). Revision count (content evolution) is weighted 2× vs duplicate count (ingest repetition). Capped at 10%.

Pin boost: Pinned documents get +0.3 additive boost (capped at 1.0). Use memory_pin to pin critical memories.

Snooze: Snoozed documents are filtered out of context surfacing until their snooze date. Use memory_snooze for temporary suppression.

Contradiction detection: When decision-extractor identifies a new decision that contradicts a prior one, the old decision's confidence is automatically lowered (−0.25 for contradictions, −0.15 for updates). Superseded decisions naturally fade from context surfacing without manual intervention. v0.7.1 adds a second layer at the consolidation boundary: before the background worker merges a new pattern into an existing consolidated observation, a deterministic heuristic (negation asymmetry, number/date mismatch) plus an LLM check detect contradictions. Blocked merges route to either link policy (both rows remain active + a contradicts edge is inserted, default) or supersede policy (old row marked status='inactive'). Configurable via CLAWMEM_CONTRADICTION_POLICY and CLAWMEM_CONTRADICTION_MIN_CONFIDENCE.

Features

A-MEM (Adaptive Memory Evolution)

Documents are automatically enriched with structured metadata when indexed:

• Keywords (3-7 specific terms)
• Tags (3-5 broad categories)
• Context (1-2 sentence description)

When new documents create links, neighboring documents' metadata evolves — keywords merge, context updates, and the evolution history is tracked with version numbers and reasoning.

Causal Inference

The decision-extractor hook analyzes Observer facts for causal relationships. When multiple facts exist in an observation, an LLM identifies cause→effect pairs (confidence ≥ 0.6). Causal chains can be queried via find_causal_links with multi-hop traversal using recursive CTEs.

Beads Integration

Projects using Beads (v0.58.0+, Dolt backend) issue tracking are fully integrated into the MAGMA memory graph:

• Auto-sync: Watcher detects .beads/ directory changes → syncBeadsIssues() queries bd CLI for live Dolt data → creates markdown docs in beads collection
• Dependency bridging: All Beads dependency types map to memory_relations edges — blocks/conditional-blocks/waits-for/caused-by→causal, discovered-from/supersedes/duplicates→supporting, relates-to/related/parent-child→semantic. Tagged {origin: "beads"} for traceability.
• A-MEM enrichment: New beads docs get full postIndexEnrich() — memory note construction, semantic/entity link generation, memory evolution
• Graph traversal: intent_search and find_causal_links traverse beads dependency edges alongside observation-inferred causal chains
• Requirement: bd binary on PATH or at ~/go/bin/bd

beads_sync MCP tool for manual sync; watcher handles routine operations automatically.

Fragment-Level Embedding

Documents are split into semantic fragments (sections, lists, code blocks, frontmatter, facts) and each fragment gets its own vector embedding. Full-doc embedding is preserved for broad-match queries.

Local Observer Agent

Uses the LLM server (shared with query expansion and intent classification) to extract structured observations from session transcripts. Observation types: decision, bugfix, feature, refactor, discovery, change, preference, milestone, problem. Each observation includes title, facts, narrative, concepts, and files read/modified. Preferences, milestones, and problems get first-class content_type treatment with dedicated confidence baselines and half-lives instead of being flattened to generic "observation". Falls back to regex patterns if the model is unavailable.

Recall Tracking

Empirical tracking of which documents are surfaced by retrieval, which queries surfaced them, and whether the assistant actually cited them. Provides signals beyond raw search relevance for lifecycle decisions:

• Per-query diversity: docs surfaced by multiple distinct queries have proven cross-domain relevance
• Multi-day spacing: docs surfaced across separate calendar days (spaced frequency) are more valuable than binge recalls in one session
• Negative signals: docs surfaced frequently but rarely referenced are noise candidates for snooze
• Per-turn attribution: feedback-loop segments the transcript into turns and attributes references to specific context-surfacing invocations, not the session globally

Data feeds lifecycle_status (pin/snooze candidate reports) and lifecycle_sweep (recall-based recommendations). Adapted from OpenClaw dreaming promotion patterns.

User Profile

Two-tier auto-curated profile extracted from your decisions and hub documents:

• Static: persistent facts (Levenshtein-deduplicated)
• Dynamic: recent session context

Injected at session start for instant personalization.

Prompt Injection Filtering

Five detection layers protect injected content: legacy string patterns, role injection regex, instruction override patterns, delimiter injection, and unicode obfuscation detection. Filtered results are skipped entirely (no placeholder tokens wasted).

Consolidation Worker

Optional background process that enriches documents missing A-MEM metadata. Runs on a configurable interval, processing 3 documents per tick. Non-blocking (Timer.unref).

Per-Folder CLAUDE.md Generation

Automatically generates context sections in per-folder CLAUDE.md files from recent decisions and session activity related to that directory.

Feedback Loop

Notes referenced by the agent during a session get boosted (access_count++). Unreferenced notes decay via recency. Over time, useful notes rise and noise fades.

Feature Flags

Variable	Default	Effect
CLAWMEM_ENABLE_AMEM	enabled	A-MEM note construction + link generation during indexing
CLAWMEM_ENABLE_CONSOLIDATION	disabled	Background worker for backlog A-MEM enrichment
CLAWMEM_CONSOLIDATION_INTERVAL	300000	Worker interval in ms (min 15000)
CLAWMEM_EMBED_URL	http://localhost:8088	Embedding server URL. Uses llama-server (GPU or CPU) or cloud API. Falls back to in-process node-llama-cpp if unset.
CLAWMEM_EMBED_API_KEY	(none)	API key for cloud embedding. Enables cloud mode: batch embedding, provider-specific params, TPM-aware pacing.
CLAWMEM_EMBED_MODEL	embedding	Model name for embedding requests. Override for cloud providers (e.g. jina-embeddings-v5-text-small).
CLAWMEM_EMBED_TPM_LIMIT	100000	Tokens-per-minute limit for cloud embedding pacing. Match to your provider tier.
CLAWMEM_EMBED_DIMENSIONS	(none)	Output dimensions for OpenAI text-embedding-3-* Matryoshka models (e.g. 512, 1024).
CLAWMEM_LLM_URL	http://localhost:8089	LLM server URL for intent/query/A-MEM. Without it, falls to node-llama-cpp (if allowed).
CLAWMEM_LLM_MODEL	qwen3	Model name sent to the configured LLM endpoint. Override this for OpenAI-compatible proxies such as gpt-5.4-mini.
CLAWMEM_LLM_REASONING_EFFORT	(none)	Optional top-level reasoning_effort field for Chat Completions endpoints that support it (for example OpenAI reasoning models). Leave unset for llama-server/vLLM unless your serving stack explicitly accepts that field.
CLAWMEM_LLM_NO_THINK	true	Append /no_think to remote LLM prompts. Set to false for standard OpenAI models and other endpoints that reject or treat the Qwen-style suffix as literal prompt text.
CLAWMEM_RERANK_URL	http://localhost:8090	Reranker server URL. Without it, falls to node-llama-cpp (if allowed).
CLAWMEM_NO_LOCAL_MODELS	false	Block node-llama-cpp from auto-downloading GGUF models. Set true for remote-only setups where you want fail-fast on unreachable endpoints.
CLAWMEM_MERGE_SCORE_NORMAL	0.93	v0.7.1. Phase 2 consolidation merge-safety threshold when candidate and existing anchors align. Merges above this normalized 3-gram cosine score are allowed.
CLAWMEM_MERGE_SCORE_STRICT	0.98	v0.7.1. Strictest merge-safety threshold — fallback when anchor sets are ambiguous.
CLAWMEM_MERGE_GUARD_DRY_RUN	false	v0.7.1. When true, Phase 2 merge-safety rejections are logged but not enforced — use for calibration before enabling the gate.
CLAWMEM_CONTRADICTION_POLICY	link	v0.7.1. Merge-time contradiction gate policy. link inserts a new row + contradicts edge (default). supersede marks the old row status='inactive'.
CLAWMEM_CONTRADICTION_MIN_CONFIDENCE	0.5	v0.7.1. Minimum combined heuristic + LLM confidence required before the contradiction gate blocks a merge.

Configuration

Collection Config

~/.config/clawmem/config.yaml:

collections:
  notes:
    path: /home/user/notes
    pattern: "**/*.md"
    autoEmbed: true
  docs:
    path: /home/user/docs
    pattern: "**/*.md"
    update: "git pull"
directoryContext: false  # opt-in per-folder CLAUDE.md generation

Database

~/.cache/clawmem/index.sqlite — single SQLite file with FTS5 + sqlite-vec extensions.

Frontmatter

Parsed via gray-matter. Supported fields:

---
title: "Document Title"
tags: [tag1, tag2]
domain: "infrastructure"
workstream: "project-name"
content_type: "decision"   # decision|preference|hub|research|project|handoff|conversation|progress|note
review_by: "2026-03-01"
---

Suggested Memory Filesystem

This structure separates human-curated content from auto-generated memories, and within auto-generated content, separates user memories (persist across agents, owned by the human) from agent memories (operational, generated from sessions). Static knowledge lives in resources/ with no recency decay, distinct from ephemeral session logs. The layout works with any MCP-compatible client (Claude Code, OpenClaw, custom agents).

Workspace Collection

The primary workspace where the agent operates. Path varies by client (e.g., ~/workspace/, ~/.openclaw/workspace/, or any directory you choose).

<workspace>/                         ← Collection: "workspace"
├── MEMORY.md                        # Human-curated long-term memory
├── memory/                          # Session logs (daily entries)
│   ├── 2026-02-05.md
│   └── 2026-02-06.md
├── resources/                       # Static knowledge — use content_type: hub (∞ half-life)
│   ├── runbooks/
│   │   └── deploy-checklist.md
│   └── onboarding.md
├── _clawmem/                        # Auto-generated — DO NOT EDIT
│   ├── user/                        #   User memories (persist across agents/sessions)
│   │   ├── profile.md               #     Static facts + dynamic context
│   │   ├── preferences/             #     Extracted preferences (update_existing merge policy)
│   │   └── entities/                #     Named entities (people, services, repos)
│   ├── agent/                       #   Agent memories (operational, session-derived)
│   │   ├── observations/            #     Decisions + observations from transcripts
│   │   ├── handoffs/                #     Session summaries with next steps
│   │   └── antipatterns/            #     Accumulated negative patterns (∞ half-life)
│   └── precompact-state.md          #   Pre-compaction snapshot (transient)
└── ...

Project Collections

Each project gets its own collection. Same structure, with optional Beads integration.

~/Projects/<project>/               ← Collection: "<project>"
├── .beads/                          # Beads issue tracker (Dolt backend, auto-synced)
│   └── dolt/                        #   Dolt SQL database (source of truth)
├── MEMORY.md                        # Human-curated project memory
├── memory/                          # Project session logs
│   └── 2026-02-06.md
├── resources/                       # Static project knowledge (∞ half-life)
│   ├── architecture.md
│   └── api-reference.md
├── research/                        # Research dumps (fragment-embedded, 90-day decay)
│   └── 2026-02-06-topic-slug.md
├── _clawmem/                        # Auto-generated per-project
│   ├── user/
│   │   └── preferences/
│   ├── agent/
│   │   ├── observations/
│   │   ├── handoffs/
│   │   ├── antipatterns/
│   │   └── beads/                   #   Beads issues as searchable markdown
│   └── precompact-state.md
├── CLAUDE.md
├── src/
└── README.md

Design Principles

Principle	Rationale	ClawMem Mechanism
User/agent separation	User memories (preferences, entities, profile) are owned by the human and persist indefinitely. Agent memories (observations, handoffs) are operational artifacts with lifecycle management.	_clawmem/user/ vs _clawmem/agent/ — different merge policies and decay rules per content type
Resources are first-class	Static knowledge (runbooks, architecture docs, API refs) should never lose relevance due to recency decay.	resources/ indexed with content_type: hub → ∞ half-life in composite scoring
Progressive disclosure	Hook injection (2000 token budget) benefits from tiered loading: compact snippets first, full content on demand.	compact=true (L1) → multi_get (L2) at query time. Pre-computed abstracts not yet implemented — candidate for future L0 tier.
Beads as memory edges	Issue tracker data bridges into the knowledge graph via typed relations, not just as flat documents.	syncBeadsIssues() maps deps → memory_relations: blocks→causal, discovered-from→supporting, relates-to→semantic
Merge policies per facet	Different memory types need different deduplication strategies to prevent bloat.	saveMemory() dedup window (30min, normalized hash) + getMergePolicy(): decision→dedup_check (cosine>0.92), antipattern→merge_recent (7d), preference→update_existing, handoff→always_new

Layer Mapping

Layer	Path	Owner	Decay	ClawMem Role
Long-Term Memory	MEMORY.md	Human	∞	Indexed, profile supplements, human-curated anchor
Session Logs	memory/*.md	Human	60 days	Indexed, daily entries, handoff auto-generated
Static Resources	resources/**/*.md	Human	∞ (hub)	Fragment-embedded, no recency penalty
Research	research/*.md	Human	90 days	Fragment-embedded for granular retrieval
User Profile	_clawmem/user/profile.md	Auto	∞	Static facts + dynamic context
User Preferences	_clawmem/user/preferences/*.md	Auto	∞	Extracted preferences (update_existing merge)
User Entities	_clawmem/user/entities/*.md	Auto	∞	Named entities across sessions
Observations	_clawmem/agent/observations/*.md	Auto	∞ (decision)	Decisions + observations from transcripts
Handoffs	_clawmem/agent/handoffs/*.md	Auto	30 days	Session summaries with next steps
Antipatterns	_clawmem/agent/antipatterns/*.md	Auto	∞	Accumulated negative patterns
Beads	_clawmem/agent/beads/*.md	Auto	∞	Beads issues synced from Dolt, relations in memory graph

Manual layers benefit from periodic re-indexing — a cron job running clawmem update --embed keeps the index fresh for content edited outside of watched directories.

Setup

# Bootstrap a content directory (creates vault + indexes + embeds + installs hooks + MCP)
clawmem bootstrap ~/notes --name notes

# Bootstrap each project you want indexed
clawmem bootstrap ~/Projects/my-project --name my-project

# Install watcher + embed timer as systemd services
clawmem install-service --enable

OpenClaw-specific

# Install the OpenClaw plugin (v0.10.0+: recursively copies into ~/.openclaw/extensions/clawmem)
clawmem setup openclaw
# Then follow the printed next steps:
#   1. openclaw plugins enable clawmem   (switches memory slot + disables memory-core)
#   2. openclaw gateway restart
#   3. Configure GPU endpoints if needed (see setup openclaw output)
# Multi-user installs also need: sudo chown -R <gateway-user>:<gateway-group> ~/.openclaw/extensions/clawmem
# Requires OpenClaw v2026.4.11+.

Index your content directories with clawmem bootstrap as above. The OpenClaw plugin shares the same vault as Claude Code hooks.

Hermes-specific

# Install the memory provider plugin.
# Preferred: $HERMES_HOME/plugins/clawmem/ (user-plugin path, Hermes #10529, v2026.4.13+).
ln -s $(npm root -g)/clawmem/src/hermes ${HERMES_HOME:-~/.hermes}/plugins/clawmem

# Bundled-style path is also supported (takes precedence on name collisions):
#   ln -s $(npm root -g)/clawmem/src/hermes /path/to/hermes-agent/plugins/memory/clawmem

# Start the REST API (required for Hermes tool calls)
clawmem serve --port 7438 &

# Configure Hermes to use ClawMem (in your Hermes config.yaml):
#   memory:
#     provider: clawmem

Dependencies

Package	Purpose
@modelcontextprotocol/sdk	MCP server
gray-matter	YAML frontmatter parsing
node-llama-cpp	GGUF model inference (reranking, query expansion, A-MEM)
sqlite-vec	Vector similarity extension
yaml	Config parsing
zod	MCP schema validation

Deployment

Three-tier retrieval architecture: infrastructure (watcher + embed timer) → hooks (~90%) → agent MCP (~10%). Works out of the box without a dedicated GPU (all models auto-download via node-llama-cpp, uses Metal on Apple Silicon). For best performance, run three llama-server instances — see GPU Services for model tiers (SOTA vs QMD native) and Cloud Embedding for cloud embedding alternatives.

Key services: clawmem-watcher (auto-index on file change + beads sync), clawmem-embed timer (daily embedding sweep), 7 Claude Code hooks installed by default (context injection, curator nudge, compaction support, decision extraction, handoffs, feedback). Optional clawmem-curator agent for on-demand lifecycle triage, retrieval health checks, and maintenance (clawmem setup curator).

Acknowledgments

Built on the shoulders of:

• A-MEM — self-evolving memory architecture
• Beads — Dolt-backed issue tracker for AI agents
• claude-mem — Claude Code memory integration reference
• Engram — observation dedup window, topic-key upsert pattern, temporal timeline navigation, duplicate metadata scoring signals
• Hermes Agent — MemoryProvider plugin integration, memory nudge system (periodic lifecycle tool prompting)
• Hindsight — entity resolution, MPFP graph traversal, temporal extraction, 3-tier consolidation, observation invalidation, 4-way parallel retrieval
• Honcho — deductive observation synthesis patterns, surprisal-based anomaly scoring concept, embed-state self-healing, retrieval separation (raw vs derived)
• MAGMA — multi-graph memory agent
• MemPalace — conversation import patterns, broadened observation taxonomy (preference/milestone/problem), session-bootstrap synthesis
• memory-lancedb-pro — retrieval gate, length normalization, MMR diversity, access reinforcement algorithms
• OpenClaw — recall tracking patterns (per-query diversity, multi-day spacing, negative signal tracking, promotion scoring) extracted from the dreaming memory consolidation system
• OpenViking — query decomposition patterns, collection-scoped retrieval, transaction-safe indexing
• QMD — search backend (BM25 + vectors + RRF + reranking)
• SAME — agent memory concepts (recency decay, confidence scoring, session tracking)
• supermemory — hook patterns and context surfacing ideas
• Thoth — anti-contamination deductive synthesis, contradiction-aware + name-aware merge gates, post-import conversation fact extraction, quiet-window heavy maintenance lane with worker leases, context instruction framing, relationship snippets, multi-turn prior-query lookback

Roadmap

Status	Feature	Description
:white_check_mark:	Adaptive thresholds	Ratio-based filtering that adapts to vault characteristics (v0.1.3)
:white_check_mark:	Deep escalation	Budget-aware query expansion + cross-encoder reranking for deep profile
:white_check_mark:	Cloud embedding providers	Jina, OpenAI, Voyage, Cohere with batch embedding + TPM pacing
:construction:	Calibration probes	One-time clawmem calibrate command that measures your vault's score distribution and tunes thresholds automatically
:construction:	Rolling threshold learning	Learns optimal activation floor from actual usage patterns (which surfaced content gets referenced vs ignored)
:memo:	Multi-vault namespacing	Isolated vaults with independent calibration and lifecycle policies
:memo:	REST API authentication	Token-scoped access for multi-agent deployments
:memo:	Streaming rerank	Cross-encoder reranking within the hook timeout budget for larger candidate sets

:white_check_mark: Shipped  :construction: Planned  :memo: Exploring

License

MIT