🧠 mnemo

Local-first agent memory CLI — dump, diff, migrate, and query memories across Mem0, Letta, and your local filesystem.

Agents are finally getting good long‑term memory, but every framework (Mem0, Letta, Supermemory, custom Postgres) stores it differently. mnemo is a git‑like CLI for agent memory: you can dump, diff, migrate, and query what your agents know, all from your terminal, using a simple normalized schema and local‑first files. It’s designed for developers who want to own their agent’s “brain” instead of locking it into a single vendor.

Inspired by Mnemosyne (Greek goddess of memory), mnemo is a portable CLI for managing agent memory: capture facts, version-control dumps, compare snapshots, and sync to cloud memory providers — all from your terminal.

Features

• 17 CLI commands with rich --help and tab-completion
• Normalized schema — facts with {entity, attribute, value, source, timestamp, confidence, metadata.tags}
• Multi-provider — local JSON, Mem0, Letta (stubs → real APIs with optional deps)
• Semantic search — mnemo recall "query" --method semantic|hybrid via fastembed (ONNX, no PyTorch); TF-IDF default needs zero extra deps
• Rich tables — confidence color-coded (🟢 ≥0.8, 🟡 ≥0.5, 🔴 <0.5)
• HTML + graph diffs — visual diff between dump snapshots
• MCP server — JSON-RPC 2.0 + stdio transport; plug directly into Claude Desktop, Cursor, or any MCP client
• Web UI — mnemo ui opens a local dashboard: browse all agents, add/edit/retract facts, import/export dumps
• Push/pull sync — S3, Cloudflare R2, or local filesystem remote; timestamp-based merge
• Auto memory from chat logs — mnemo ingest --file chat.json extracts facts from Claude.ai, ChatGPT, Cursor, or plain text exports via Claude API, OpenAI, Ollama, or heuristics
• Python SDK — MnemoClient + AsyncMnemoClient for programmatic access; local file I/O or remote HTTP; file-watch stream; no CLI required
• Safe writes — --dry-run on load, pull, migrate, and ingest

Quick Start

# Install
pip install mnemo-agent                    # core (local only)
pip install "mnemo-agent[semantic]"        # + semantic/hybrid search (fastembed, no PyTorch)
pip install "mnemo-agent[ingest]"          # + auto memory from chat logs (anthropic + openai SDKs)
pip install "mnemo-agent[s3]"             # + S3/R2 push-pull sync
pip install "mnemo-agent[sdk]"            # + Python SDK (MnemoClient, AsyncMnemoClient — httpx)
pip install "mnemo-agent[all]"            # everything (mem0 + letta + parquet + graph + s3 + semantic + ingest + sdk)

# Initialize Joshua's job-prep agent
mnemo init --agent job-prep

# Add facts manually (entity defaults to agent name, --tag is repeatable)
mnemo add --fact "Joshua uses React, Node, Supabase, Vercel" --agent job-prep
mnemo add --fact "Joshua is based in Toronto" --agent job-prep --confidence 1.0
mnemo add --fact "Chose Supabase over Firebase for auth" --agent job-prep --attribute decision --tag decision --tag auth

# View stored memories — plain format shows IDs for retract/edit
mnemo show --agent job-prep
mnemo show --agent job-prep --format plain

# Recall using natural language, optionally filtered by tag
mnemo recall "tech stack" --agent job-prep                          # TF-IDF (default)
mnemo recall "tech decisions I've made" --agent job-prep --method semantic   # fastembed semantic
mnemo recall "auth" --agent job-prep --tag decision --method hybrid          # hybrid
mnemo search "Supabase database" --agent job-prep --limit 5

# Auto-extract facts from a Claude.ai or ChatGPT chat export
mnemo ingest --file ~/Downloads/claude_chat.json --agent job-prep
mnemo ingest --file ~/Downloads/conversations.json --format chatgpt --agent job-prep
mnemo ingest --file chat.json --agent job-prep --extractor heuristic  # offline, no API key
mnemo ingest --file chat.json --agent job-prep --dry-run              # preview before saving

# Edit or remove facts by ID (use 'show --format plain' to find IDs)
mnemo retract a1b2c3d4 --agent job-prep
mnemo edit a1b2c3d4 --value "Updated wording" --agent job-prep

# List all agents
mnemo ls --pretty

# Dump to a timestamped file
mnemo dump --agent job-prep

# Load a sample dump
mnemo load --file tests/fixtures/job_prep_sample.json --agent job-prep

# Compare two agents (or two dump files)
mnemo diff --agent-a job-prep --agent-b job-prep-v2
mnemo diff dump1.json dump2.json --html diff_report.html

# Start the MCP server — HTTP mode
mnemo serve --agent job-prep --port 8080
# Or stdio mode (Claude Desktop / Cursor — no port needed)
mnemo serve --agent job-prep --stdio

# Open the web dashboard (all agents, auto-opens browser)
mnemo ui
# Or jump straight to a specific agent
mnemo ui --agent job-prep

# Sync to S3 (prompts for credentials on first add)
mnemo remote add origin s3://my-bucket/mnemo --agent job-prep
mnemo push --agent job-prep
mnemo pull --agent job-prep   # merges remote facts into local

🔍 Semantic Search

mnemo supports three search modes via --method:

pip install "mnemo-agent[semantic]"   # ~30MB install, model (~130MB) downloads on first use

mnemo recall "tech decisions I've made" --method semantic
mnemo recall "outdoor activities" --method hybrid --limit 10

• Model: BAAI/bge-small-en-v1.5 (384-dim, cached at ~/.cache/fastembed/ after first use)
• Speed: ~0.1–0.5s for <500 facts on CPU — fast enough for CLI use
• MCP: pass "mode": "semantic" or "mode": "hybrid" in search_memory tool args
• REST: GET /search?q=query&mode=semantic or GET /agents/{agent}/search?q=query&mode=hybrid

🤖 Auto Memory from Chat Logs

mnemo ingest parses a chat export and uses an LLM (or heuristics offline) to extract structured facts — no manual add required.

pip install "mnemo-agent[ingest]"   # anthropic + openai SDKs (~10MB combined)

# Auto-picks best extractor: ANTHROPIC_API_KEY → OPENAI_API_KEY → Ollama → heuristic
mnemo ingest --file ~/Downloads/claude_chat.json --agent job-prep

# Explicit extractor + model
mnemo ingest --file chat.json --agent job-prep --extractor openai --extractor-model gpt-4o

# Ollama (local, fully offline)
mnemo ingest --file chat.json --agent job-prep --extractor ollama --extractor-model llama3.2

# Gemini via OpenAI-compatible endpoint
OPENAI_API_KEY=$GEMINI_KEY mnemo ingest --file chat.json --agent job-prep \
  --extractor openai \
  --extractor-url https://generativelanguage.googleapis.com/v1beta/openai/ \
  --extractor-model gemini-1.5-flash

# Heuristic (regex patterns, zero deps)
mnemo ingest --file chat.json --agent job-prep --extractor heuristic

Always shows a preview table and asks Save these N facts? [y/N] before writing. Use --dry-run to preview without the prompt.

Supported formats: Claude.ai JSON, ChatGPT conversations.json, Cursor, plain text (auto-detected).

🐍 Python SDK

Use mnemo programmatically — no CLI required. Works locally (reads ~/.mnemo directly) or remotely (talks to mnemo serve).

pip install "mnemo-agent[sdk]"   # adds httpx for remote mode; local mode needs nothing extra

from mnemo import MnemoClient, AsyncMnemoClient

# ── Sync local (reads ~/.mnemo directly) ──────────────────────────────────
client = MnemoClient(agent="job-prep")

fact = client.add("Lead with Supabase story", attribute="interview_tip", confidence=0.9)
results = client.recall("Supabase", method="hybrid")   # tfidf | semantic | hybrid
facts   = client.list_facts(attribute="interview_tip", tag="tip")
client.edit(fact.id[:8], value="Updated wording")
client.retract(fact.id[:8])

info = client.info()   # {"agent": "job-prep", "fact_count": 12, "last_updated": "..."}
dump = client.dump()   # returns AgentDump

# ── Sync remote (talks to mnemo serve) ───────────────────────────────────
remote = MnemoClient(agent="job-prep", url="http://localhost:8080")
remote.add("Rust for systems work", attribute="tool")

# ── Async (local or remote) ───────────────────────────────────────────────
import asyncio

async def main():
    async with AsyncMnemoClient(agent="job-prep") as client:
        fact  = await client.add("async fact", attribute="note")
        found = await client.recall("fact", method="tfidf")
        print(found)

asyncio.run(main())

# ── File-watch stream (local only) ────────────────────────────────────────
async def watch_demo():
    async with AsyncMnemoClient(agent="job-prep") as client:
        async for fact in client.watch(poll_interval=0.5):
            print(f"New: {fact.entity} · {fact.attribute}: {fact.value}")

SDK method reference

AsyncMnemoClient mirrors every method as async def. Both clients work as context managers.

📋 All Commands

Project Structure

mnemo-agent/
├── src/mnemo/
│   ├── __init__.py          # version + SDK exports (MnemoClient, AsyncMnemoClient, Fact, AgentDump)
│   ├── cli.py               # Click CLI (all commands)
│   ├── client.py            # Python SDK: MnemoClient + AsyncMnemoClient (local + remote backends)
│   ├── models.py            # Pydantic: Fact, AgentDump, MnemoConfig
│   ├── storage.py           # Local file I/O (JSON, YAML, credentials)
│   ├── search.py            # TF-IDF, semantic, and hybrid search + diff engine
│   ├── embeddings.py        # fastembed backend for semantic search (optional dep)
│   ├── remotes.py           # Push/pull backends: FileBackend, S3Backend
│   ├── server.py            # FastAPI MCP server (HTTP + JSON-RPC 2.0) + multi-agent UI server
│   ├── stdio_server.py      # stdio MCP transport (Claude Desktop / Cursor)
│   ├── static/
│   │   └── ui.html          # Single-file web dashboard (Alpine.js + Tailwind CDN)
│   └── adapters/
│       ├── mem0_adapter.py     # Mem0 API → normalized facts
│       ├── letta_adapter.py    # Letta API → normalized facts
│       └── ingest_adapter.py   # Chat log parsers + LLM/heuristic fact extractors
├── tests/
│   ├── test_cli.py          # CLI command tests
│   ├── test_client.py       # SDK tests: local, remote (mocked), async, file-watch
│   ├── test_ingest.py       # Chat parsers, heuristic extractor, ingest CLI tests
│   ├── test_remote.py       # Remote backends, merge, push/pull tests
│   ├── test_server.py       # MCP server: JSON-RPC 2.0, tools, stdio transport
│   └── fixtures/
│       └── job_prep_sample.json
├── config.yaml              # Sample agent config
├── pyproject.toml
└── requirements.txt

Memory Schema

{
  "agent": "job-prep",
  "dump_ts": "2026-03-21T23:00Z",
  "source": "manual",
  "version": "1",
  "facts": [
    {
      "id": "uuid",
      "entity": "Joshua",
      "attribute": "tech_stack",
      "value": "React, Node, Supabase, Vercel",
      "source": "chat|tool|manual|mem0|letta|import",
      "timestamp": "2026-03-21T20:00Z",
      "confidence": 0.95,
      "metadata": {}
    }
  ]
}

Example: Project Memory for advisor-prep

{
  "agent": "advisor-prep",
  "facts": [
    {
      "id": "uuid-1",
      "entity": "advisor-prep-agent",
      "attribute": "project_summary",
      "value": "CLI + agent that helps students prep for advisor meetings using UBC context.",
      "source": "manual",
      "timestamp": "2026-03-22T01:00Z",
      "confidence": 0.9,
      "metadata": { "tags": ["summary", "high-level"] }
    },
    {
      "id": "uuid-2",
      "entity": "advisor-prep-agent",
      "attribute": "decision",
      "value": "Chose Supabase over Firebase for auth due to better Postgres integration.",
      "source": "manual",
      "timestamp": "2026-03-22T01:05Z",
      "confidence": 0.95,
      "metadata": { "tags": ["decision", "auth"], "ticket": "ADR-001" }
    },
    {
      "id": "uuid-3",
      "entity": "advisor-prep-agent",
      "attribute": "stack",
      "value": "Next.js, React, Node, Supabase, Vercel.",
      "source": "manual",
      "timestamp": "2026-03-22T01:10Z",
      "confidence": 1.0,
      "metadata": { "tags": ["stack"] }
    }
  ]
}

🔌 MCP Server

mnemo implements the MCP 2024-11-05 spec and supports two transports.

stdio — Claude Desktop / Cursor

Add to ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "mnemo-job-prep": {
      "command": "mnemo",
      "args": ["serve", "--agent", "job-prep", "--stdio"]
    }
  }
}

That's it — Claude Desktop will spawn mnemo as a subprocess and communicate over stdin/stdout.

HTTP — REST + JSON-RPC 2.0

mnemo serve --agent job-prep --port 8080

Available MCP tools

retract_fact and edit_fact are disabled when --read-only is set.

🖥 Web Dashboard

mnemo ui                        # opens http://localhost:7742/ui
mnemo ui --agent job-prep       # deep-links to that agent
mnemo ui --port 8080 --read-only

mnemo ui requires uvicorn (pip install uvicorn). The browser opens automatically.

Agent list view

• Cards for every agent — fact count, dump count, last updated, top tags
• Create a new agent directly from the UI
• Delete an agent (confirmation required)
• Click any card to open the agent detail view

Agent detail view

• Facts table — entity, attribute, value, confidence bar, tags, relative age
• Filter chips — one-click entity/attribute filters above the table; tag filter in sidebar
• Add / Edit / Retract facts with a slide-in panel
• Import — upload a dump JSON, merges new facts by ID
• Export — download the agent's latest dump as <agent>-dump.json
• Search — TF-IDF results with relevance scores (semantic/hybrid available via MCP/CLI)
• Diff — upload a second dump file and see added/removed/unchanged facts side by side

⚙️ Configuration

Each agent has ~/.mnemo/<agent>/config.yaml:

agent: job-prep
default_source: local
default_target: local
mem0_api_key: null          # https://app.mem0.ai
mem0_user_id: joshua
letta_base_url: http://localhost:8283
letta_agent_id: null        # from your Letta agent
tags: [job-prep, interview]
notes: Memory store for interview prep agent
remotes:
  origin: s3://my-bucket/mnemo

Remote credentials (S3/R2 access keys) are stored separately in ~/.mnemo/credentials with chmod 600. They are populated automatically when you run mnemo remote add — you will be prompted for them interactively. Pass --no-creds to skip prompting and rely on the standard boto3 credential chain (AWS_ACCESS_KEY_ID env var, ~/.aws/credentials, IAM role).

Environment Variables

Tests

pip install "mnemo-agent[dev,s3]"
pytest tests/ -v

# Include semantic search tests
pip install "mnemo-agent[dev,s3,semantic]"
pytest tests/ -v

187+ tests across test_cli.py, test_client.py, test_ingest.py, test_remote.py, and test_server.py. Semantic search tests are auto-skipped when mnemo[semantic] is not installed. Ingest tests mock all external API calls. SDK remote tests mock _get/_mcp_call to avoid live HTTP.

The web UI (mnemo ui) is served by the same FastAPI process as mnemo serve and is covered by the existing server tests.

Roadmap

• Push/pull sync to S3, R2, and local filesystem remotes
• Write-time conflict detection with overwrite / keep-both / abort prompt
• Full MCP 2024-11-05 protocol — JSON-RPC 2.0 + stdio transport (Claude Desktop / Cursor)
• Web UI dashboard — multi-agent overview, per-agent facts/search/diff/import/export
• Semantic search — --method semantic|hybrid via fastembed (ONNX, no PyTorch); hybrid combines TF-IDF + cosine with configurable alpha
• Auto memory from chat logs — mnemo ingest extracts facts from Claude.ai/ChatGPT/Cursor/plain exports via Claude API, OpenAI-compatible (Groq, Gemini, Ollama), or offline heuristics
• Python SDK — MnemoClient + AsyncMnemoClient with local and remote backends; watch() file-stream; pip install mnemo-agent[sdk]
• Parquet export for analytics
• mnemo audit — fact provenance trace
• Snapshot history browser in UI

Example Use Case: job-prep Agent

# Bootstrap your interview prep memory
mnemo init --agent job-prep
mnemo load --file tests/fixtures/job_prep_sample.json --agent job-prep

# Connect to Claude Desktop (add to claude_desktop_config.json, then restart)
mnemo serve --agent job-prep --stdio

# Or run as an HTTP server for other MCP clients
mnemo serve --agent job-prep --port 8080

# After a practice interview, add what you learned
mnemo add --fact "Lead with Supabase migration story at FAANG interviews" \
  --agent job-prep --attribute interview_tip --confidence 0.9 --tag tip

# If you added a conflicting fact by mistake, retract it by ID prefix
mnemo show --agent job-prep --format plain   # see IDs
mnemo retract a1b2c3d4 --agent job-prep

# Before next session, recall relevant context
mnemo recall "React Supabase full-stack" --agent job-prep

License

MIT © Joshua Ndala