Remnic
Local-first memory plugin for OpenClaw AI agents. LLM-powered extraction, plain markdown storage, hybrid search via QMD. Gives agents persistent long-term memory across conversations.
Ask AI about Remnic
Powered by Claude Β· Grounded in docs
I know everything about Remnic. Ask me about installation, configuration, usage, or troubleshooting.
0/500
Reviews
Documentation
Remnic
Persistent, private memory for AI agents. Your agents forget everything between sessions β Remnic fixes that.
The trace is noise. The primitive is the product. Remnic's job is the pipeline that distills hours of agent conversation into compressed, searchable, durable memory primitives. (How it works β)
Remnic gives AI agents long-term memory that survives across conversations. Decisions, preferences, project context, personal details, past mistakes β everything your agent learns persists and resurfaces exactly when it's needed. All data stays on your machine as plain markdown files. No cloud services, no subscriptions, no sharing your data with third parties.
Engram is now Remnic. Canonical packages live under the
@remnic/*scope:@remnic/core,@remnic/server,@remnic/cli. OpenClaw installs should use@remnic/plugin-openclaw. The legacyengramCLI name remains available as a forwarder during the rename window. Hermes users:remnic-hermesv1.0.2 on PyPI.
Support Remnic
Every bit of support is genuinely appreciated and helps keep this project alive and free for everyone.
If you're able to, sponsoring on GitHub or sending a Lightning donation to joshuaswarren@strike.me directly funds continued development, new integrations, and keeping Remnic open source.
If financial support isn't an option, you can still make a big difference β star the repo on GitHub, share it on social media, or recommend it to a friend or colleague. Word of mouth is how most people find Remnic, and it means the world.
The Problem
Every AI agent session starts from zero. Your agent doesn't know your name, your projects, the decisions you've already made, or the bugs you already debugged. Whether it's a personal assistant, a coding agent, a research agent, or a multi-agent team β they all forget everything between conversations. You re-explain the same context over and over, and your agents still make the same mistakes.
OpenClaw's built-in memory works for simple cases, but it doesn't scale. It lacks semantic search, lifecycle management, entity tracking, and governance. Third-party memory services exist, but they cost money and require sending your private data to someone else's servers.
The Solution
Remnic is an open-source, local-first memory system that replaces OpenClaw's default memory with something much more capable β while keeping everything on your machine. It watches your agent conversations, extracts durable knowledge, and injects the right memories back at the start of every session. Route extraction through the OpenClaw gateway model chain, OpenAI, or a local LLM (Ollama, LM Studio, etc.) β your choice.
Remnic is the universal memory layer for AI agents. It works natively with OpenClaw, Claude Code, Codex CLI, Hermes Agent, and any MCP-compatible client (Replit, Cursor, etc.). When you tell any agent a preference, every agent knows it β they share one memory store.
Architecture rule: standalone Remnic is first-class. @remnic/core, @remnic/server, and @remnic/cli own the memory engine and must stay host-agnostic. OpenClaw, Hermes, Codex, Claude Code, and future integrations are thin adapters over that shared core, and adapter work should always follow the host's current upstream SDK and documentation instead of recreating host-native behavior inside Remnic.
| Without Remnic | With Remnic |
|---|---|
| Re-explain who you are and what you're working on | Agent recalls your identity, projects, and preferences automatically |
| Repeat context for every task | Entity knowledge surfaces people, projects, tools, and relationships on demand |
| Lose debugging and research context between sessions | Past root causes, dead ends, and findings are recalled β no repeated work |
| Manually restate preferences every session | Preferences persist across sessions, agents, and projects |
| Context-switching tax when resuming work | Session-start recall brings you back to speed instantly |
| Default OpenClaw memory doesn't scale | Hybrid search, lifecycle management, namespaces, and governance |
| Third-party memory services cost money and share your data | Everything stays local β your filesystem, your rules |
Memory or context substrate? Both.
There's a useful split in the AI-memory space between memory backends (extract facts β vector DB β retrieve relevant ones) and context substrates (structured human-readable context that accumulates across sessions and compounds over time). Most tools land firmly in one camp. Remnic does both.
The files are the source of truth. Every memory is a markdown file with YAML frontmatter on your filesystem. You can cat, grep, edit, version-control, back up, and reason about your memory with standard tools. The hybrid search index (QMD: BM25 + vector + reranking) is downstream of the files β fully rebuildable from disk, never the source of truth itself.
The recall stays sharp. Three retrieval tiers (chunk β section β raw transcript), feature-flagged graph retrieval with Personalized PageRank, memory-worth scoring that filters low-value facts before they reach the LLM, temporal supersession that keeps stale facts out of recall, and Recall X-ray so you can see exactly which tier produced each result and why.
It compounds. Background consolidation (the "dreams" surface) merges duplicates, promotes recurring themes, and snapshots page versions on every overwrite. Provenance fields (derived_from, derived_via) track where every consolidated memory came from. Procedural memory (on by default) captures multi-step runbooks. The longer you use it, the better it gets β and you can always read exactly what it knows.
Camp 1 asks "what should the AI remember?" Remnic answers that. Camp 2 asks "what context should the AI work inside?" Remnic answers that too.
Quick install (OpenClaw)
If you have OpenClaw installed, the fastest path to working Remnic memory is:
# 1. Install the plugin package
openclaw plugins install clawhub:@remnic/plugin-openclaw
# 2. Wire up the memory slot automatically
remnic openclaw install
# 3. Restart the gateway
launchctl kickstart -k gui/$(id -u)/ai.openclaw.gateway
# 4. Verify everything is working
remnic doctor
remnic openclaw install writes plugins.entries["openclaw-remnic"] and plugins.slots.memory = "openclaw-remnic" to ~/.openclaw/openclaw.json. Without the slot, hooks never fire β see Troubleshooting: hooks aren't firing for details.
Migrating from the legacy @joshuaswarren/openclaw-engram package? Run
remnic openclaw migrate-engram --yes; it backs up the legacy extension,
installs @remnic/plugin-openclaw, preserves memoryDir, and switches the
OpenClaw memory slot to openclaw-remnic. See the
OpenClaw Engram to Remnic migration guide.
Installation
Option 1: Install from the CLI
openclaw plugins install clawhub:@remnic/plugin-openclaw
Option 2: Ask your OpenClaw agent to install it
Tell any OpenClaw agent:
Install the @remnic/plugin-openclaw plugin and configure it as my memory system.
Your agent will run the install command, update openclaw.json, and restart the gateway for you.
Option 3: Developer install from source
git clone https://github.com/joshuaswarren/remnic.git \
~/.openclaw/extensions/remnic
cd ~/.openclaw/extensions/remnic
pnpm install && pnpm run build
Note: This repo uses pnpm workspaces.
npm ci/npm installwill fail onworkspace:specifiers. Install pnpm first:npm install -g pnpm.
Option 4: Standalone (no OpenClaw)
From npm (recommended):
npm install -g @remnic/cli # Installs `remnic` plus the legacy `engram` forwarder
remnic init # Create remnic.config.json
export OPENAI_API_KEY=sk-...
export REMNIC_AUTH_TOKEN=$(openssl rand -hex 32)
remnic daemon start # Start background server
remnic status # Verify it's running
remnic query "hello" --explain # Test query with tier breakdown
From source (requires Node.js 22.12+ and pnpm):
git clone https://github.com/joshuaswarren/remnic.git
cd remnic
pnpm install && pnpm run build
cd packages/remnic-cli && pnpm link --global # Makes `remnic` and `engram` available on PATH
cd ../..
remnic init # Create remnic.config.json
export OPENAI_API_KEY=sk-...
export REMNIC_AUTH_TOKEN=$(openssl rand -hex 32)
remnic daemon start # Start background server
remnic status # Verify it's running
remnic query "hello" --explain # Test query with tier breakdown
Note:
remnicis the canonical CLI. The legacyengrambinary is a compatibility forwarder to the same implementation. Runningpnpm link --globalfrompackages/remnic-cli/(not the repo root) makes both names available on PATH. Alternatively, invoke directly:npx tsx packages/remnic-cli/src/index.ts <command>.
The standalone CLI provides 15+ commands for memory management, project onboarding, curation, diff-aware sync, dedup, connectors, spaces, and benchmarks -- all without requiring OpenClaw. See the Platform Migration Guide for the full command reference.
Option 5: Connect Other AI Agents
Once the Remnic daemon is running, connect any supported agent:
remnic connectors install claude-code # Claude Code (hooks + MCP)
remnic connectors install codex-cli # Codex CLI (hooks + MCP + memory extension)
remnic connectors install replit # Replit (MCP only)
pip install --upgrade remnic-hermes # Hermes Agent (Python MemoryProvider)
remnic connectors install hermes # Writes Hermes config + token
For Codex CLI, installation also drops a phase-2 memory extension at
<codex_home>/memories_extensions/remnic/instructions.md so Codex's
consolidation sub-agent auto-discovers Remnic. Opt out with
--config installExtension=false if you prefer to manage Codex extensions
yourself.
Each connector generates a unique auth token, installs the appropriate plugin/hooks, and verifies the connection. All agents share the same memory store β tell one agent your preference, and every agent remembers it.
Hermes uses Remnic as a Hermes MemoryProvider, not a context_engine. Automatic recall runs in pre_llm_call, observations run after each turn, and the provider now registers the full Remnic parity tool surface (remnic_lcm_search, recall explain/X-ray, memory CRUD, continuity, identity, governance, work-board, shared-context, compounding, day-summary, briefing, checkpoint, and profiling tools) plus legacy engram_* aliases. Lossless Context Management is delivered through the daemon recall envelope when lcmEnabled is on; no Hermes context_engine registration is required. See docs/plugins/hermes.md for the full reference.
| Platform | Integration | Auto-recall | Auto-observe |
|---|---|---|---|
| OpenClaw | Memory slot plugin | Every session | Every response |
| Claude Code | Native hooks + MCP | Every prompt | Every tool use |
| Codex CLI | Native hooks + MCP | Every prompt | Every tool use |
| Hermes | Python MemoryProvider | Every LLM call | Every turn |
| Replit | MCP only | On demand | On demand |
Configure
After installation, add the Remnic bridge plugin to your openclaw.json:
{
"plugins": {
"allow": ["openclaw-remnic"],
"slots": { "memory": "openclaw-remnic" },
"entries": {
"openclaw-remnic": {
"enabled": true,
"hooks": {
"allowConversationAccess": true
},
"config": {
// Recommended for OpenClaw: use the gateway model chain.
"modelSource": "gateway",
"gatewayAgentId": "remnic-llm",
"fastGatewayAgentId": "remnic-llm-fast",
// Optional: Use Remnic's local LLM path (plugin mode only; no API key needed):
// "localLlmEnabled": true,
// "localLlmUrl": "http://localhost:1234/v1",
// "localLlmModel": "qwen2.5-32b-instruct"
// Optional: Use OpenAI directly (plugin mode only):
// "modelSource": "plugin",
// "openaiApiKey": "${OPENAI_API_KEY}"
}
}
}
}
}
Gateway model source: When
modelSourceis"gateway", Remnic routes all LLM calls (extraction, consolidation, reranking) through an OpenClaw agent persona's model chain instead of its own config. Extraction starts on thegatewayAgentIdchain directly in this mode;localLlm*settings do not control primary extraction order. Define agent personas inopenclaw.json β agents.list[]with aprimarymodel andfallbacks[]array β Remnic tries each in order until one succeeds. This lets you build multi-provider fallback chains like Fireworks β local LLM β cloud OpenAI. See the Gateway Model Source guide for full setup.
Restart the gateway:
launchctl kickstart -k gui/$(id -u)/ai.openclaw.gateway # macOS
# or: systemctl restart openclaw-gateway # Linux
Start a conversation β Remnic begins learning immediately.
Note: This shows only the minimal config. Remnic has 60+ configuration options for search backends, capture modes, memory OS features, and more. See the full config reference for every setting.
Extraction importance gate
Remnic scores every extracted fact locally (see src/importance.ts) and uses that score as a write gate. Facts whose level falls below extractionMinImportanceLevel are dropped before they ever hit disk, so turn-level chatter like "hi", "k", or heartbeat pings never become fact memories.
Default: "low" β only "trivial" content is dropped. Raise to "normal" or higher for a stricter gate.
{
"plugins": {
"entries": {
"openclaw-remnic": {
"config": {
// Allowed values: "trivial" | "low" | "normal" | "high" | "critical"
"extractionMinImportanceLevel": "normal"
}
}
}
}
}
Category boosts still apply before the gate, so corrections, principles, preferences, and commitments stay above "normal" even when their raw text would otherwise score low. Every gated fact increments the importance_gated counter (grep metric:importance_gated in ~/.openclaw/logs/gateway.log) and the final extraction log line reports the gated count.
Inline source attribution (opt-in, issue #369)
Extracted facts can optionally carry a compact provenance tag inline in the fact body β not just in YAML frontmatter β so the citation survives prompt injection, copy/paste, and LLM quoting. When an agent later quotes a memory back or a user asks "where did you learn that?", the source travels with the claim.
Default format:
The foo service uses Redis for rate limiting. [Source: agent=planner, session=main, ts=2026-04-10T14:25:07Z]
Enable it per plugin:
{
"plugins": {
"entries": {
"openclaw-remnic": {
"config": {
"inlineSourceAttributionEnabled": true,
// Optional: customize the tag format.
// Placeholders: {agent}, {session}, {sessionId}, {ts}, {date}
"inlineSourceAttributionFormat": "[Source: agent={agent}, session={sessionId}, ts={ts}]"
}
}
}
}
}
Properties:
- Off by default to preserve backwards compatibility with downstream consumers that expect raw fact text.
- Inline β the tag is part of the stored fact body, so it flows through every write site (direct writes, chunked writes, shared-namespace promotion, verbatim artifacts) and recall injection without special handling.
- Legacy-safe β facts written before the flag was enabled still read and recall normally; nothing is retroactively rewritten.
- Non-destructive β facts that already carry a citation (e.g. relayed from an upstream system) are left untouched.
- Machine-parseable β
parseCitation(text)andstripCitation(text)are exported from@remnic/corefor callers that want the raw body (e.g. for dedup hashing, display, or verification tooling). Malformed citations never throw.
See packages/remnic-core/src/source-attribution.ts for the helpers and packages/remnic-core/src/source-attribution.test.ts for the round-trip contract.
Verify installation
remnic doctor # Health diagnostics with remediation hints
remnic connectors doctor # Connector-specific health checks
remnic status # Daemon status and local endpoint summary
Bring your memory
Remnic can import existing memory from the platforms you already use. Five optional importer packages ship alongside the CLI β install only the ones you need:
# ChatGPT (OpenAI data export: saved memories + optional conversation summaries)
npm install -g @remnic/import-chatgpt
remnic import --adapter chatgpt --file ~/chatgpt-export/memory.json --dry-run
# Claude (Anthropic data export: project docs + prompt templates)
npm install -g @remnic/import-claude
remnic import --adapter claude --file ~/claude-export/projects.json
# Gemini (Google Takeout "Gemini Apps Activity")
npm install -g @remnic/import-gemini
remnic import --adapter gemini --file "~/Takeout/Gemini/My Activity.json"
# mem0 (REST API β paginated; honors --rate-limit)
npm install -g @remnic/import-mem0
export MEM0_API_KEY=...
remnic import --adapter mem0 --rate-limit 2
# Supermemory (JSON export)
npm install -g @remnic/import-supermemory
remnic import --adapter supermemory --file ./supermemory-memories.json --dry-run
remnic import --adapter supermemory --file ./supermemory-memories.json
Each importer is an optional runtime companion β the base CLI
install never pulls them in. If you run remnic import --adapter <name>
without the matching package installed, the CLI prints a clean install
hint. Every run supports --dry-run for a zero-write preview.
Privacy note: import parsing and storage run locally, but after the orchestrator accepts a record it enters the normal extraction pipeline β which calls whatever model provider you have configured. If extraction is routed to a remote provider, imported content is transmitted to that provider during extraction. To keep imports fully local, configure a local extraction model or use
--dry-runto preview without writing.
See docs/importers.md for per-source details, input formats, provenance metadata, and the full privacy breakdown.
Troubleshooting: hooks aren't firing
Symptom: Remnic appears installed but no memories are created. The gateway log shows no [remnic] lines after conversations.
Root cause: OpenClaw gates memory plugins on plugins.slots.memory. If this slot is not set to the plugin's id, OpenClaw skips register(api) entirely β no hooks fire, no memory is stored or recalled.
Quick fix
remnic openclaw install # Sets plugins.slots.memory = "openclaw-remnic"
Restart the gateway after running this command.
How to verify hooks are firing
After restarting, check the gateway log for this line:
[remnic] gateway_start fired β Remnic memory plugin is active (id=openclaw-remnic, memoryDir=~/.openclaw/workspace/memory/local)
On macOS:
grep "gateway_start fired" ~/.openclaw/logs/gateway.log
If the line is absent, run remnic doctor to see which check is failing:
remnic doctor
The doctor output will show:
OpenClaw config fileβ whetheropenclaw.jsonexists and is valid JSONOpenClaw plugins.entriesβ whether the entries object is presentOpenClaw plugin entryβ whetheropenclaw-remnic(or legacyopenclaw-engram) entry existsOpenClaw plugins.slots.memoryβ whether the slot is set and points to an entryOpenClaw memoryDirβ whether the configured memory directory exists on disk
Each failing check includes a remediation hint pointing to remnic openclaw install.
Manual fix
If you prefer to edit ~/.openclaw/openclaw.json directly:
{
"plugins": {
"entries": {
"openclaw-remnic": {
"config": {
"memoryDir": "~/.openclaw/workspace/memory/local"
}
}
},
"slots": {
"memory": "openclaw-remnic"
}
}
}
Both entries["openclaw-remnic"] and slots.memory = "openclaw-remnic" are required. See docs/integration/plugin-id-and-memory-namespaces.md for the full design note.
Using Remnic with Codex CLI
Start the Remnic server directly for the current shell session:
# Generate a token
export REMNIC_AUTH_TOKEN="$(openssl rand -base64 32)"
npx remnic-server --host 127.0.0.1 --port 4318 --auth-token "$REMNIC_AUTH_TOKEN"
If you want to use remnic daemon start, persist the token in
remnic.config.json first. daemon start will hand off to launchd/systemd
when a service is installed, and those service templates read server.authToken
from config rather than inheriting your shell's exported token.
The HTTP API path remains /engram/v1/... during the v1.x compatibility window.
Add to ~/.codex/config.toml:
[mcp_servers.remnic]
url = "http://127.0.0.1:4318/mcp"
bearer_token_env_var = "REMNIC_AUTH_TOKEN"
That's it. Codex now has access to Remnic's recall, store, and entity tools. See the full Codex integration guide for session-start hooks, cross-machine setup, and automatic recall at session start.
Using Remnic with Any MCP Client
Run the stdio MCP server:
openclaw engram access mcp-serve
Point your MCP client's command at openclaw engram access mcp-serve. This
is the OpenClaw-hosted stdio compatibility path. For standalone Remnic installs,
prefer the HTTP MCP endpoint exposed by remnic daemon start or remnic-server.
Claude Code (MCP over HTTP): Start the Remnic server, then add to ~/.claude.json:
{
"mcpServers": {
"remnic": {
"url": "http://localhost:4318/mcp",
"headers": {
"Authorization": "Bearer ${REMNIC_AUTH_TOKEN}"
}
}
}
}
See the Standalone Server Guide for multi-tenant setups and connecting multiple agent harnesses.
Standalone Usage
Remnic also works as a standalone tool without OpenClaw. Install and run the CLI directly:
npm install -g @remnic/cli
remnic init # create remnic.config.json
export OPENAI_API_KEY=sk-...
export REMNIC_AUTH_TOKEN=$(openssl rand -hex 32)
remnic daemon start # start background server
remnic query "hello" # verify
The CLI provides 15+ commands for querying, onboarding projects, curating files, managing spaces, running benchmarks, and more. See the full CLI reference for all commands.
Connect to any coding tool
Remnic works with 10+ coding tools via MCP or HTTP. See the Connector Setup Guide for config snippets for Claude Code, Codex CLI, Cursor, GitHub Copilot, Cline, Roo Code, Windsurf, Amp, Replit, and any generic MCP client.
OpenClaw remains the recommended path for most users. The standalone CLI is useful for CI/CD pipelines, scripted memory operations, and environments without OpenClaw.
Package Architecture
@remnic/core β Framework-agnostic engine (re-exports orchestrator, config, storage, search, extraction, graph, trust zones)
@remnic/cli β Standalone CLI binary (15+ commands)
@remnic/server β Standalone HTTP/MCP server
@remnic/bench β Benchmarks + CI regression gates
@remnic/hermes-provider β HTTP client for remote Remnic instances
How It Works
Remnic operates in three phases:
Recall β Before each conversation, inject relevant memories into context
Buffer β After each turn, accumulate content until a trigger fires
Extract β Periodically, extract structured memories using an LLM
Memories are stored as plain markdown files with YAML frontmatter β fully portable, git-friendly, no database required:
---
id: decision-1738789200000-a1b2
category: decision
confidence: 0.92
tags: ["architecture", "search"]
---
Decided to use the port/adapter pattern for search backends
so alternative engines can replace QMD without changing core logic.
Memory categories include: fact, decision, preference, correction, relationship, principle, commitment, moment, skill, rule, and more.
Architecture
Remnic is organized as a monorepo with a core engine, standalone server/CLI, and native plugins for multiple AI platforms:
βββββββββββββββββββ
β @remnic/core β
β (engine) β
ββββββββββ¬βββββββββ
β
ββββββββββββ¬βββββββββββ¬ββββ΄βββββ¬βββββββββββ¬βββββββββββ
β β β β β β
βββββββ΄ββββββ βββ΄βββββββ ββ΄ββββββ ββ΄βββββββββ β Native β
β @remnic/ β β@remnic/β βremnicβ β@remnic/ β β Plugins β
β cli β βserver β β-hermesβ βplugin- β β β
β β β β β β βopenclaw β β β
βββββββββββββ ββββββββββ ββββββββ βββββββββββ ββββββββββββ
β β
βββββββ΄ββββββ ββββββββββββββββΌβββββββββββ
β @remnic/ β β β β
β bench β claude-code codex replit
βββββββββββββ
Remnic is installed Γ la carte: most users start with @remnic/cli, while @remnic/core stays available for framework-agnostic embedding. Optional surfaces (bench, WeClone, plugins, and importers) are installed separately when you need them. Commands like remnic bench *, remnic training:export, and remnic import --adapter <source> lazy-load their companion package and print an install hint if it's missing.
The old @joshuaswarren/openclaw-engram package is deprecated. Use @remnic/plugin-openclaw for OpenClaw installs and @remnic/* for standalone or multi-platform use.
Why Remnic?
Your data stays yours
All memory lives on your filesystem as plain markdown files. No cloud dependency, no subscriptions, no proprietary formats, no sending your private conversations to third-party servers. Back it up with git, rsync, or Time Machine. Move it between machines with a folder copy. You own your data completely.
A real upgrade from default OpenClaw memory
OpenClaw's built-in memory is basic β it works for getting started, but lacks semantic search, entity tracking, lifecycle management, governance, and multi-agent isolation. Remnic is a drop-in replacement that brings all of those capabilities while keeping the same local-first philosophy.
Smart recall, not keyword search
Remnic uses hybrid search (BM25 + vector + reranking via QMD) to find semantically relevant memories. It doesn't just match keywords β it understands what you're working on and surfaces the right context.
Flexible LLM routing β OpenAI, local, or gateway model chain
Use OpenAI for extraction and reranking, run entirely offline with a local LLM (Ollama, LM Studio), or route through the gateway model chain to use any provider with automatic fallback. The local-llm-heavy preset is optimized for fully local operation. See the Local LLM Guide and the Gateway Model Source section for multi-provider setups.
Progressive complexity
Start with zero config. Enable features as your needs grow:
| Level | What You Get |
|---|---|
| Defaults | Automatic extraction, recall injection, entity tracking, lifecycle management |
| + Search tuning | Choose from 6 search backends (QMD, Orama, LanceDB, Meilisearch, remote, noop) |
| + Capture control | implicit, explicit, or hybrid capture modes for memory write policy |
| + Memory OS | Memory boxes, graph reasoning, compounding, shared context, identity continuity |
| + LCM | Lossless Context Management β never lose conversation context to compaction |
| + Parallel retrieval | Three specialized agents (DirectFact, Contextual, Temporal) run in parallel β same latency, broader coverage |
| + Quality gates | Extraction judge, semantic chunking, MECE taxonomy, page versioning |
| + Advanced | Trust zones, causal trajectories, harmonic retrieval, evaluation harness, poisoning defense |
Use a preset to jump to a recommended level: conservative, balanced, research-max, or local-llm-heavy.
Works with your tools
- OpenClaw β Native plugin with automatic extraction and recall injection
- Codex CLI β MCP-over-HTTP with session-start hooks for automatic recall
- Any MCP client β stdio or HTTP transport, 8 tools available
- Scripts & automation β Authenticated REST API for custom integrations
- Local LLMs β Run extraction and reranking with local models (Ollama, LM Studio, etc.)
Standalone Multi-Tenant Server
Run Remnic as a standalone HTTP server that multiple agent harnesses share. Isolate tenants with namespace policies, feed conversations from any client via the observe endpoint, and search archived history with LCM full-text search. Works with OpenClaw, Codex CLI, Claude Code, and custom HTTP agents. See the Standalone Server Guide.
Built for production
- 672 tests with CI enforcement
- Evaluation harness with benchmark packs, shadow recall recording, and CI delta gates
- Governance system with review queues, shadow/apply modes, and reversible transitions
- Namespace isolation for multi-agent deployments
- Rate limiting on write paths with idempotency support
Features
Core
- Automatic memory extraction β Facts, decisions, preferences, corrections extracted from conversations
- Observe endpoint β Feed conversation messages from any agent into the extraction pipeline via HTTP or MCP
- Recall injection β Relevant memories injected before each agent turn
- Entity tracking β People, projects, tools, companies tracked as structured entities
- Lifecycle management β Memories age through active, validated, stale, archived states
- Episode/Note model β Memories classified as time-specific events or stable beliefs
Extraction & Processing (opt-in)
- Extraction Judge β LLM-as-judge post-extraction filter that evaluates fact durability before write. Has shadow mode for calibration. Opt-in via
extractionJudgeEnabled. (issue #376) - Semantic Chunking β Smoothing-based topic boundary detection using sentence embeddings and cosine similarity, as an alternative to recursive chunking. Opt-in via
semanticChunkingEnabled. (issue #368) - OAI-mem-citation Blocks β Recall responses emit
<oai-mem-citation>blocks matching the Codex citation format for memory attribution and usage tracking. Opt-in viacitationsEnabled. (issue #379) - Procedural memory β Stores repeatable how-to memories as
category: proceduremarkdown underprocedures/, mines candidates from causal trajectories, and can inject a Relevant procedures section on task-initiation prompts. On by default since issue #567 PR 4/5 (previously off); setprocedural.enabledtofalsein plugin config to opt out. See Procedural memory. (issue #519) - Peer registry β Multi-peer identity schema that generalizes the singular identity-anchor into a versioned registry. Supports
self,human,agent, andintegrationpeer kinds. Each peer has an identity kernel (peers/{id}/identity.md), an async profile reasoner that derives structured fields from interaction signals, and recall injection that injects a brief profile excerpt into recall context.remnic peer migrateseedspeers/self/identity.mdfrom existing legacy identity-anchor data.remnic peer list/show/set/delete/profilemanage the registry. HTTP endpoints and MCP tools underpeer_*. See Peer Registry. (issue #679) - Coding agent mode β Auto-scopes memory to the current git project (origin-URL hash) and optionally to the current branch, so memories from project A never surface in project B and feature-branch experiments don't leak into
main. Claude Code and Codex CLI session-start hooks detect git context automatically; therecallandobserveendpoints also acceptcwdfor server-side auto-resolution. Non-git sessions (OpenClaw, task agents) can passprojectTagto scope by name instead. Cross-project knowledge (framework bugs, user preferences) is classified as"global"during extraction and promoted to the shared namespace; recall global fallback ensures it surfaces across all projects (codingMode.globalFallback, defaulttrue). Diff-aware review-context recall tier boosts memories touching the files in a reviewed diff.remnic doctorprints detectedprojectId, branch, and effective namespace. Opt out withcodingMode.projectScope: false. See Coding agent mode. (issues #569, #702, #703, #704) - Recall X-ray β
remnic xray "<query>"prints a per-result breakdown showing which retrieval tier served each memory, the score decomposition, the graph path (when graph retrieval fired), the filter ladder that admitted it, and the audit entry ID. Same snapshot via HTTPGET /engram/v1/recall/xrayand MCP toolremnic.recall_xray. Legacy/recall/explaingains a markdown mode that delegates to the same renderer (backwards-compatible). See Recall X-ray. (issue #570) - Disclosure depth on recall β
remnic recall "<q>" --disclosure chunk|section|rawcontrols payload depth: cheap semantic chunks first, escalate to full sections or raw transcripts only when needed. Same field exposed via HTTP?disclosure=and MCPdisclosure. Defaultchunkpreserves prior behavior; token-cost telemetry surfaces in Recall X-ray. (issue #677) - Temporal recall (
validAt/invalidAt) β OptionalvalidAtandinvalidAtYAML frontmatter fields scope a fact to a validity window;remnic recall "<q>" --as-of <timestamp>returns the fact valid at that time, ignoring later supersessions. Supersession writes flipinvalidAtautomatically. Backfill defaultsvalidAtto the file'screatedtimestamp when missing. See Temporal Recall. (issue #680) - Free-form tag filter on recall β
remnic recall "<q>" --tag q2-planning [--tag <t> ...] [--tag-match all|any]. Tags are flat strings additive to the MECE taxonomy, persisted on memory frontmatter via the store API. Tags surface in Recall X-ray output. See Tags. (issue #689)
Organization & Taxonomy (opt-in)
- MECE Taxonomy β Mutually Exclusive, Collectively Exhaustive knowledge directory with resolver decision tree for deterministic memory categorization. Opt-in via
taxonomyEnabled. (issue #366) - Enrichment Pipeline β Importance-tiered API spend for entity enrichment from external sources with a provider registry. Opt-in via
enrichmentEnabled. (issue #365)
Storage & Lifecycle (opt-in)
- Page Versioning β Snapshot-based history for memory files. Every overwrite saves a numbered snapshot in a sidecar directory. List, inspect, diff, and revert. Opt-in via
versioningEnabled. (issue #371) - Binary Lifecycle Management β Three-stage pipeline (mirror, redirect, clean) for binary files in the memory directory with configurable storage backends. Opt-in via
binaryLifecycleEnabled. (issue #367)
Integrations & Extensions
- Codex Marketplace β Install Remnic via
codex marketplace add joshuaswarren/remnic. Marketplace manifest at repo root. (issue #418) - Memory Extension Publisher Contract β Pluggable contract for installing host-specific instruction files into any AI agent host's extension directory. Generalizes the pattern previously hard-coded for Codex. (issue #381)
- Memory Extension Discovery β Third-party memory extensions provide structured instructions that influence consolidation, auto-discovered from extension directories. (issue #382)
Search Backends
| Backend | Type | Best For |
|---|---|---|
| QMD (default) | Hybrid BM25+vector+reranking | Best recall quality |
| Orama | Embedded, pure JS | Zero native deps |
| LanceDB | Embedded, native Arrow | Large collections |
| Meilisearch | Server-based | Shared search |
| Remote | HTTP REST | Custom services |
| Noop | No-op | Extraction only |
See the Search Backends Guide or write your own.
Memory OS (opt-in)
These capabilities can be enabled progressively:
- Memory Boxes β Groups related memories into topic-windowed episodes
- Graph Recall β Entity-relationship graph for causal and timeline queries
- Compounding β Weekly synthesis surfaces patterns and recurring mistakes
- Shared Context β Cross-agent memory sharing for multi-agent setups
- Identity Continuity β Consistent agent personality across sessions
- Hot/Cold Tiering β Automatic migration of aging memories to cold storage
- Memory Cache β Process-level singleton cache for
readAllMemories()β turns 15s disk scans into <100ms cache hits, shared across all sessions - Semantic Consolidation β Finds clusters of semantically similar memories, synthesizes canonical versions via LLM, archives originals to reduce bloat
- Native Knowledge β Search curated markdown (workspace docs, Obsidian vaults) without extracting into memory
- Behavior Loop Tuning β Runtime self-tuning of extraction and recall parameters
Lossless Context Management (LCM)
When your AI agent hits its context window limit, the runtime silently compresses old messages β and that context is gone forever. LCM fixes this by proactively archiving every message into a local SQLite database and building a hierarchical summary DAG (directed acyclic graph) alongside it. When context gets compacted, LCM injects compressed session history back into recall, so your agent never loses track of what happened earlier in the conversation.
- Proactive archiving β Every message is indexed with full-text search before compaction can discard it
- Hierarchical summaries β Leaf summaries cover ~8 turns, depth-1 covers ~32, depth-2 ~128, etc.
- Fresh tail protection β Recent turns always use the most detailed (leaf-level) summaries
- Three-level summarization β Normal LLM summary, aggressive bullet compression, and deterministic truncation (guaranteed convergence, no LLM needed)
- MCP expansion tools β Agents can search, describe, or expand any part of conversation history on demand
- Zero data loss β Raw messages are retained for the configured retention period (default 90 days)
Enable it in your openclaw.json:
{
"plugins": {
"entries": {
"openclaw-remnic": {
"config": {
"lcmEnabled": true
// All other LCM settings have sensible defaults
}
}
}
}
}
See the LCM Guide for architecture details, configuration options, and how it complements native compaction.
Parallel Specialized Retrieval (opt-in)
Remnic's default retrieval runs a single hybrid search pass. Parallel Specialized Retrieval (inspired by Supermemory's ASMR technique) runs three specialized agents in parallel so total latency equals max(agents) not sum(agents).
| Agent | What It Does | Cost |
|---|---|---|
| DirectFact | Scans entity filenames for keyword overlap with the query | File I/O only, <5ms |
| Contextual | Existing hybrid BM25+vector search (unchanged) | Same as current |
| Temporal | Reads the temporal date index, returns recent memories with recency decay scoring | File I/O + math, <10ms |
Zero additional LLM cost. The DirectFact and Temporal agents reuse existing indexes with no new embeddings or inference. The Contextual agent is the same hybrid search already running.
Results from all three agents are merged by path, deduplicated, and weighted (direct=1.0Γ, temporal=0.85Γ, contextual=0.7Γ) before returning the top N results. Any agent error degrades gracefully without blocking the others.
Enable it in your openclaw.json:
{
"plugins": {
"entries": {
"openclaw-remnic": {
"config": {
"parallelRetrievalEnabled": true
// Optional tuning:
// "parallelMaxResultsPerAgent": 20,
// "parallelAgentWeights": { "direct": 1.0, "contextual": 0.7, "temporal": 0.85 }
}
}
}
}
}
Set parallelMaxResultsPerAgent: 0 to disable an individual agent's results without disabling the feature entirely.
Semantic Consolidation (opt-in)
Over time, memory stores accumulate redundant facts β the same information extracted multiple times across sessions, expressed slightly differently. Semantic consolidation finds clusters of similar memories using token overlap, synthesizes a single canonical version via LLM, and archives the originals. This reduces storage bloat, speeds up recall, and improves memory quality.
- Conservative by default β Only merges when 80%+ token overlap is detected across 3+ memories
- LLM synthesis β Uses your configured model to combine unique information from all cluster members
- Safe archival β Originals are archived (not deleted) with full provenance tracking
- Configurable β Adjust threshold, cluster size, excluded categories, model, and schedule
- Excluded categories β Corrections and commitments are never consolidated (configurable)
Enable it in your openclaw.json:
{
"plugins": {
"entries": {
"openclaw-remnic": {
"config": {
"semanticConsolidationEnabled": true
// Optional tuning:
// "semanticConsolidationThreshold": 0.8, // 0.8=conservative, 0.6=aggressive
// "semanticConsolidationModel": "fast", // "auto", "fast", or specific model
// "semanticConsolidationIntervalHours": 168, // weekly (default)
// "semanticConsolidationMaxPerRun": 100
}
}
}
}
}
Run manually from the CLI:
openclaw engram semantic-consolidate --dry-run # Preview what would be merged
openclaw engram semantic-consolidate --verbose # Run with detailed output
openclaw engram semantic-consolidate --threshold 0.6 # Override threshold
Advanced (opt-in)
- Objective-State Recall β Surfaces file/process/tool state snapshots alongside semantic memory
- Causal Trajectories β Typed
goal -> action -> observation -> outcomechains - Trust Zones β Quarantine/working/trusted tiers with promotion rules and poisoning defense
- Harmonic Retrieval β Blends abstraction nodes with cue-anchor matches
- Verified Recall β Only surfaces memory boxes whose source memories still verify
- Semantic Rule Promotion β Promotes
IF ... THENrules from verified episodes - Creation Memory β Work-product ledger tracking agent outputs
- Commitment Lifecycle β Tracks promises, deadlines, and obligations
- Resume Bundles β Crash-recovery context for interrupted sessions
- Utility Learning β Learns promotion/ranking weights from downstream outcomes
See Enable All Features for a full-feature config profile.
Access Layer
Remnic exposes one shared service layer through multiple transports. During the
v1.x compatibility window, the HTTP API path remains /engram/v1/... and the
legacy engram.* MCP aliases still work.
HTTP API
remnic daemon start
Key endpoints: GET /engram/v1/health, POST /engram/v1/recall, POST /engram/v1/memories, GET /engram/v1/entities/:name, and more. Full reference in API docs.
The HTTP server also hosts a lightweight operator UI at http://127.0.0.1:4318/engram/ui/ for memory browsing, recall inspection, governance review, trust-zone promotion, and entity exploration.
MCP Tools
Available via both stdio and HTTP transports:
| Tool | Purpose |
|---|---|
engram.recall | Retrieve relevant memories for a query |
engram.recall_explain | Debug the last recall |
engram.day_summary | Generate structured end-of-day summary from memory content |
engram.memory_get | Fetch a specific memory by ID |
engram.memory_timeline | View a memory's lifecycle history |
engram.memory_store | Store a new memory |
engram.suggestion_submit | Queue a memory for review |
engram.entity_get | Look up a known entity |
engram.review_queue_list | View the governance review queue |
engram.observe | Feed conversation messages into memory pipeline (LCM + extraction) |
engram.lcm_search | Full-text search over LCM-archived conversations |
engram_context_search | Full-text search across all archived conversation history (LCM) |
engram_context_describe | Get a compressed summary of a turn range (LCM) |
engram_context_expand | Retrieve raw lossless messages for a turn range (LCM) |
MCP over HTTP
The HTTP server exposes an MCP JSON-RPC endpoint at POST /mcp, allowing remote MCP clients to use Remnic tools over HTTP:
npx remnic-server --host 0.0.0.0 --port 4318 --auth-token "$REMNIC_AUTH_TOKEN"
For namespace-enabled deployments, configure server.principal in remnic.config.json so it matches a writePrincipals entry for your target namespace. Deployments with namespacesEnabled: false (the default) do not need a principal.
CLI Reference
# OpenClaw-hosted compatibility commands still use the `openclaw engram`
# namespace during the v1.x rename window. Standalone commands use `remnic`.
#
# Setup & diagnostics
openclaw engram setup # Guided first-run setup
openclaw engram doctor # Health diagnostics with remediation hints
openclaw engram config-review # Config tuning recommendations
openclaw engram stats # Memory counts, search status
openclaw engram inventory # Full storage and namespace inventory
# Search & recall
openclaw engram search "query" # Search memories from CLI
openclaw engram harmonic-search "query" # Preview harmonic retrieval matches
# Governance
openclaw engram governance-run --mode shadow # Preview governance transitions
openclaw engram governance-run --mode apply # Apply reversible transitions
openclaw engram review-disposition <id> --status rejected # Operator review
# Benchmarking
openclaw engram benchmark recall # Benchmark status and validation
openclaw engram benchmark-ci-gate # CI gate for regressions
# Memory maintenance
openclaw engram consolidate # Run standard consolidation
openclaw engram semantic-consolidate # Run semantic dedup consolidation
openclaw engram semantic-consolidate --dry-run # Preview without changes
# Daily context briefing (#370)
remnic briefing # Yesterday's briefing (markdown)
remnic briefing --since 3d --focus project:alpha # Focused 3-day lookback
remnic briefing --format json --save # JSON + dated file in $REMNIC_HOME/briefings
# Page versioning
remnic versions list <page-path> # List version history for a memory file
remnic versions show <page-path> <version-id> # Show a specific version snapshot
remnic versions diff <page-path> <v1> <v2> # Diff two versions of a memory file
remnic versions revert <page-path> <version-id> # Revert a file to a previous version
# MECE taxonomy
remnic taxonomy show # Show taxonomy categories and priorities
remnic taxonomy resolver # Generate or display resolver decision tree
remnic taxonomy add <id> <name> # Add a taxonomy category
remnic taxonomy remove <id> # Remove a taxonomy category
# Entity enrichment
remnic enrich <entity-name|--all|audit|providers> [--dry-run] # Run enrichment pipeline
# Binary lifecycle
remnic binary scan # Scan for binary files in memory directory
remnic binary status # Show binary lifecycle status
remnic binary run [--dry-run] # Run lifecycle (redirect/clean) for binaries
remnic binary clean --force # Force-clean binaries past grace period
# Access layer
remnic daemon start # Start HTTP API + managed daemon
openclaw engram access mcp-serve # Start OpenClaw-hosted stdio MCP server
# Trust-zone demos
openclaw engram trust-zone-demo-seed --dry-run # Preview the opt-in buyer demo dataset
openclaw engram trust-zone-demo-seed # Explicitly seed the demo dataset
openclaw engram trust-zone-promote --record-id <id> --target-zone working --reason "Operator review"
Trust-zone demo workflow
Trust zones now ship with a dedicated admin-console view plus an explicit demo seeding path for buyer-facing walkthroughs.
- Never automatic β Remnic does not seed sample trust-zone records on install, startup, or feature enablement.
- Explicit only β demo records appear only after you run
openclaw engram trust-zone-demo-seedor trigger the matching admin-console action. - Buyer-friendly story β the trust-zone view surfaces provenance strength, promotion readiness, corroboration requirements, and operator promotion actions in one place.
The seeded scenario is enterprise-buyer-v1, which creates a small, opinionated dataset covering:
- quarantine records that are ready for review
- working records that are blocked on missing provenance
- working records that still need corroboration
- working records with independent corroboration support
- a trusted operator policy record
See the full CLI reference for all commands.
Configuration
OpenClaw plugin settings live in openclaw.json under plugins.entries.openclaw-remnic.config (with a legacy openclaw-engram fallback during the rename window). Standalone settings live in remnic.config.json or ~/.config/remnic/config.json. The table below shows the most commonly changed settings β Remnic has 60+ configuration options covering search backends, capture modes, memory OS features, namespaces, governance, benchmarking, and more.
| Setting | Default | Description |
|---|---|---|
modelSource | gateway for new OpenClaw installs; plugin otherwise | gateway routes LLM calls through OpenClaw agent model chains; plugin uses Remnic's own OpenAI/local LLM config |
openaiApiKey | (env in plugin mode) | Optional OpenAI API key. Not needed when modelSource is gateway; Remnic does not inherit OPENAI_API_KEY in gateway mode. |
localLlmEnabled | false | Enable Remnic's local LLM path when modelSource is plugin |
localLlmUrl | unset | Local LLM endpoint (e.g., http://localhost:1234/v1) |
localLlmModel | unset | Local model name (e.g., qwen2.5-32b-instruct) |
model | gpt-5.2 | OpenAI model for extraction when modelSource is plugin and local LLM is disabled |
searchBackend | "qmd" | Search engine: qmd, orama, lancedb, meilisearch, remote, noop |
captureMode | implicit | Memory write policy: implicit, explicit, hybrid |
recallBudgetChars | maxMemoryTokens * 4 | Recall budget (default ~8K chars; set 64K+ for large-context models) |
memoryDir | ~/.openclaw/workspace/memory/local | Memory storage root |
memoryOsPreset | unset | Quick config: conservative, balanced, research-max, local-llm-heavy |
lcmEnabled | false | Enable Lossless Context Management (proactive session archive + summary DAG) |
messagePartsEnabled | false | Opt in to structured LCM message-part capture for tool calls, file paths, patches, and reasoning markers |
messagePartsRecallMaxResults | 6 | Max structured file/tool matches injected into recall when messagePartsEnabled is on |
semanticConsolidationEnabled | false | Enable periodic semantic dedup of similar memories |
semanticConsolidationThreshold | 0.8 | Token overlap threshold (0.8=conservative, 0.6=aggressive) |
semanticConsolidationModel | "auto" | LLM for synthesis: "auto", "fast", or specific model |
extractionJudgeEnabled | false | LLM-as-judge post-extraction durability filter |
semanticChunkingEnabled | false | Topic-boundary chunking via sentence embeddings |
versioningEnabled | false | Snapshot-based page versioning with history and revert |
citationsEnabled | false | Emit oai-mem-citation blocks in recall responses |
taxonomyEnabled | false | MECE knowledge directory with resolver decision tree |
enrichmentEnabled | false | External entity enrichment pipeline |
binaryLifecycleEnabled | false | Binary file lifecycle management (mirror/redirect/clean) |
procedural.enabled | true | Procedural memory (issue #519): master gate for procedure extraction, task-init recall injection, and trajectory mining. Default-on since issue #567 PR 4/5; set nested procedural: { "enabled": false } to opt out (see Procedural memory). |
codingMode.projectScope | true | Coding agent mode (issue #569): auto-scope memory to the git project (stable origin-URL hash, falls back to root path). Set to false to disable project-based namespace isolation. See Coding agent mode. |
codingMode.branchScope | false | Coding agent mode (issue #569): additionally scope writes to the current git branch; reads fall back to the project-level namespace so project memories stay visible from any branch while branch writes don't leak. Enable for per-branch experimentation. |
codingMode.globalFallback | true | Recall global fallback (issue #703): project-scoped sessions include the root/global namespace in recall read-fallbacks so that cross-project knowledge (framework bugs, library behavior, user preferences) surfaces everywhere. Set to false for strict project isolation. See Coding agent mode. |
extractionScopeClassificationEnabled | true | Extraction scope classification (issue #704): classify extracted facts as "global" or "project" scope. Global facts are promoted to the shared root namespace so they are visible across all projects. See Coding agent mode. |
recallCrossNamespaceBudgetEnabled | false | Cross-namespace budget (issue #565): per-principal sliding-window rate limiter. Throttles principals issuing bursts of cross-namespace recalls; soft limit emits a warning, hard limit denies the query. See Threat model. |
recallAuditAnomalyDetectionEnabled | false | Recall audit anomaly detection (issue #565): flag suspicious query patterns (repeat queries, namespace walks, high-cardinality entity probes, rapid-fire). Anomalies are surfaced in recall responses. See Threat model. |
connectors.googleDrive.enabled | false | Google Drive live connector (issue #683 PR 2/N): opt-in incremental import of Google Docs/Sheets/Slides + plain-text files. Requires clientId, clientSecret, refreshToken to be populated from a secret store (never commit values). Optional pollIntervalMs (default 300000) and folderIds (default [] = all accessible). The googleapis npm package is loaded lazily β install it only if you enable the connector. See Connectors. |
connectors.notion.enabled | false | Notion live connector (issue #683 PR 3/N): opt-in incremental import of Notion database pages. Requires token (Notion integration token) and databaseIds (list of database IDs to import). Optional pollIntervalMs (default 300000). See Connectors. |
graphTraversalConfidenceFloor | 0.2 | Graph edge confidence floor (issue #681 PR 3/3): minimum edge confidence required during graph spreading activation. Edges below this floor are pruned and contribute neither activation nor downstream neighbors. Legacy edges without a confidence field are treated as 1.0. Range [0, 1]. See Graph Reasoning. |
graphTraversalPageRankIterations | 8 | Graph PageRank refinement (issue #681 PR 3/3): number of PageRank-style refinement iterations applied on top of the BFS spreading-activation scores. Each iteration redistributes a node's confidence-weighted activation along its outgoing edges. Set to 0 to disable refinement and use raw BFS scores. See Graph Reasoning. |
patternReinforcementEnabled | false | Pattern reinforcement (issue #687): master gate for the cross-session pattern-reinforcement maintenance job. Default false (opt-in). See Pattern Reinforcement. |
patternReinforcementCadenceMs | 604800000 | Minimum milliseconds between pattern-reinforcement runs (default 7 days). Set to 0 to disable cadence gating and run on every MCP/cron invocation. |
patternReinforcementMinCount | 3 | Minimum cluster size required before a canonical memory is reinforced. Clamped to [2, 1000]. |
patternReinforcementCategories | ["preference", "fact", "decision"] | Memory categories the pattern-reinforcement job scans. Set to [] to process no categories. |
recallDisclosureEscalation | "manual" | Recall disclosure auto-escalation (issue #677): "manual" honors the caller's requested disclosure depth verbatim. "auto" escalates from chunk to section when top-K confidence falls below recallDisclosureEscalationThreshold β only on calls where the caller did not specify a depth. raw is never auto-selected. See Recall Disclosure. |
recallDisclosureEscalationThreshold | 0.5 | Disclosure escalation threshold (issue #677): confidence threshold in [0, 1] used by recallDisclosureEscalation: "auto". Recalls whose top-result score is below this value are escalated from chunk to section. Has no effect in manual mode. See Recall Disclosure. |
reinforcementRecallBoostEnabled | false | Reinforcement recall boost (issue #687 PR 3/4): when true, memories whose reinforcement_count frontmatter field is set receive an additive score boost during recall. Default false (opt-in). |
reinforcementRecallBoostWeight | 0.05 | Score bonus per unit of reinforcement_count. Raw boost is weight Γ count, clipped at reinforcementRecallBoostMax. Range [0, 1]. |
reinforcementRecallBoostMax | 0.3 | Maximum additive reinforcement boost per result. Range [0, 1]. |
See the full config reference for all 60+ settings including search backend configuration, namespace policies, Memory OS features, governance, evaluation harness, trust zones, causal trajectories, and more.
Documentation
- Getting Started β Installation, setup, first-run verification
- Config Reference β Every setting with defaults
- Architecture Overview β System design and storage layout
- Retrieval Pipeline β How recall works
- Memory Lifecycle β Write, consolidation, expiry
- Search Backends β Choosing and configuring search engines
- Writing a Search Backend β Build your own adapter
- API Reference β HTTP, MCP, and CLI documentation
- Codex CLI Integration β Set up Remnic with OpenAI's Codex
- Standalone Server Guide β Multi-tenant HTTP server for multiple agent harnesses
- Local LLM Guide β Local-first extraction and reranking
- Cost Control Guide β Budget mappings and presets
- Namespaces β Multi-agent memory isolation
- Shared Context β Cross-agent intelligence
- Identity Continuity β Consistent agent personality
- Peer Registry β Multi-peer identity schema, profile reasoner,
remnic peercommands, and migration from legacy identity-anchor (issue #679) - Graph Reasoning β Opt-in graph traversal
- Evaluation Harness β Benchmarks and CI delta gates
- Operations β Backup, export, maintenance
- Lossless Context Management β Never lose context to compaction
- Enable All Features β Full-feature config profile
- Migration Guide β Upgrading from older versions
- Platform Migration Guide β Migrating to the monorepo architecture (v9.1.36+)
- Hermes Setup β HTTP client for remote Remnic instances
- Deployment Topologies β Localhost, LAN, remote, containerized, standalone
- Extraction Judge β LLM-as-judge fact-worthiness gate
- Semantic Chunking β Topic-boundary detection
- Page Versioning β Snapshot-based history and revert
- Citations β OAI-mem-citation block format
- Memory Extension Publishers β Pluggable publisher contract
- MECE Taxonomy β Knowledge directory with resolver
- Enrichment Pipeline β Entity enrichment from external sources
- Binary Lifecycle β Binary file management
- Memory Extensions β Third-party extension discovery
- Codex Marketplace β Marketplace installation
- Procedural memory β Procedure files, recall injection, mining; enable with
procedural.enabled(issue #519) - Pattern Reinforcement β Cross-session pattern detection, reinforced primitives,
remnic patterns list/explainCLI, recall boost (issue #687) - Coding agent mode β Auto-scope memory to git project / branch, review-context recall,
set_coding_contextMCP tool (issue #569) - Recall X-ray β
remnic xrayCLI, HTTP endpoint, MCP tool for per-result retrieval attribution (issue #570) - Connectors β
remnic connectors list/status/runCLI reference, OAuth setup, config keys, env vars, and troubleshooting for Google Drive and Notion (issue #683 PR 6/N) - Live connectors framework β Connector framework contract, registry, state store API, and how to write a connector
- Memory importers β Bring memory from ChatGPT, Claude, Gemini, mem0, and Supermemory (issue #568)
- Memory Extraction Threat Model β ADAM attack analysis, attacker tiers, and mitigation wiring (issue #565)
- ADAM Baseline 2026-04 β Reproducible ASR measurements per attacker tier
Contributing
Contributions are welcome! Please:
- Fork the repository
- Create a feature branch (
git checkout -b feat/my-feature) - Write tests for new functionality
- Ensure
npm test(672 tests) andnpm run check-typespass - Submit a pull request
