Persistent, private memory for MCP agents. Runs entirely on your hardware.
No cloud. No API keys. No subscriptions.

Prerequisites: Python 3.11+ and Ollama (or any OpenAI-compatible embedding endpoint).

ollama pull nomic-embed-text && ollama serve
pip install m3-memory

Add to your MCP config (~/.claude/settings.json, ~/.gemini/settings.json, etc.):

{
  "mcpServers": {
    "memory": { "command": "mcp-memory" }
  }
}

Restart your agent. It now has memory. Done.

Already inside Claude Code, Gemini CLI, or Aider? Paste one of these prompts and let the agent set it up for you:

Claude Code:

Install m3-memory for persistent memory. Run: pip install m3-memory
Then add {"mcpServers":{"memory":{"command":"mcp-memory"}}} to my
~/.claude/settings.json under "mcpServers". Make sure Ollama is running
with nomic-embed-text. Then use /mcp to verify the memory server loaded.

Gemini CLI:

Install m3-memory for persistent memory. Run: pip install m3-memory
Then add {"mcpServers":{"memory":{"command":"mcp-memory"}}} to my
~/.gemini/settings.json under "mcpServers". Make sure Ollama is running
with nomic-embed-text.

Aider / Any MCP agent:

Install m3-memory for persistent memory. Run: pip install m3-memory
Then add {"mcpServers":{"memory":{"command":"mcp-memory"}}} to the
MCP config file for this agent. Make sure Ollama is running with
nomic-embed-text.

After install, test it:

Write a memory: "M3 Memory installed successfully on [today's date]"
Then search for: "M3 install"

Every new session, your AI agent has amnesia. It forgets your project structure, your preferences, the decisions you made together yesterday. You paste the same context. You re-explain the same architecture. You correct the same mistakes.

When facts change — a port number, a dependency version — there's no mechanism to update what the agent "knows." Contradictions accumulate silently until something breaks.

Your agents remember. Architecture decisions, server configs, debugging history, your preferences — all searchable, all persistent across sessions and devices.

When facts change, M3 detects the contradiction, updates the record, and preserves the full history. No stale data. No manual cleanup. You just talk to your agent, and it knows what it should know.

Without M3: "I don't have that information. Could you tell me?"

With M3: "Port 9000. (Updated from 8080 — change recorded March 12th.)"

No prompts. No manual logic. Automatic contradiction resolution with full history.

Three-stage pipeline: FTS5 keyword matching, semantic vector similarity, and MMR diversity re-ranking. Better recall than vector-only search, especially for technical content with exact names and versions.

Write conflicting information and M3 detects it automatically. The outdated memory is superseded via bitemporal versioning, a supersedes relationship is recorded, and the full history is preserved.

Query as_of="2026-01-15" to see exactly what your agent believed on any past date. Every change is tracked with both the time the fact was true and the time it was recorded.

Related facts are linked on write when cosine similarity exceeds 0.7. Eight relationship types (related, supports, contradicts, extends, supersedes, references, consolidates, message). Traverse up to 3 hops with memory_graph.

Bi-directional delta sync across SQLite, PostgreSQL, and ChromaDB. Write on your MacBook, continue on your Windows desktop. No cloud intermediary.

gdpr_forget (Article 17 — Right to Erasure) and gdpr_export (Article 20 — Data Portability) as native MCP tools.

Local embeddings via Ollama, LM Studio, or any OpenAI-compatible endpoint. Zero cloud calls. Zero API costs. Works completely offline.

Automatic decay, expiry purging, orphan pruning, deduplication, and retention enforcement. Old memories consolidate into LLM-generated summaries.

Start with three — memory_write, memory_search, and memory_update — that covers 90% of daily use. The rest is there when you need it.

Full list of all 25 tools

graph TD
    subgraph "AI Agents"
        C[Claude Code]
        G[Gemini CLI]
        A[Aider / OpenClaw]
    end

    subgraph "MCP Bridge"
        MB[memory_bridge.py — 25 MCP tools]
    end

    subgraph "Storage Layers"
        SQ[(SQLite — Local L1)]
        PG[(PostgreSQL — Sync L2)]
        CH[(ChromaDB — Federated L3)]
    end

    C & G & A <--> MB
    MB <--> SQ
    SQ <-->|Bi-directional Delta Sync| PG
    SQ <-->|Push/Pull| CH

sequenceDiagram
    participant A as Agent
    participant M as M3 Memory
    participant L as Local LLM
    participant S as SQLite

    A->>M: memory_write(content)
    M->>M: Safety Check (XSS / injection / poisoning)
    M->>L: Generate Embedding
    L-->>M: Vector [0.12, -0.05, ...]
    M->>M: Contradiction Detection
    M->>M: Auto-Link Related Memories
    M->>M: SHA-256 Content Hash
    M->>S: Store Memory + Vector
    S-->>M: Success
    M-->>A: Created: <uuid>

Agent writes two conflicting facts — the old one is automatically superseded:

FTS5 keyword + semantic vector + MMR diversity re-ranking — with full explainability:

Write a memory on your MacBook, search it on your Windows desktop — no cloud:

Have a real recording to share? See CONTRIBUTING.md or post in #showcase on Discord.

Get help, share your setup, and follow development. M3_Bot is live — use !ask <question> in any channel.

Vote on features in ROADMAP.md

bin/          MCP bridge, core engine, sync, and maintenance scripts
m3_memory/    Python package — CLI entry point (mcp-memory)
memory/       SQLite database and migrations
docs/         Architecture diagrams and install guides
examples/     Demo notebooks and ready-to-paste mcp.json configs
tests/        End-to-end test suite (41 tests)

1. Star the repo — helps others find it
2. Try a real session — install, write a memory, close your agent, reopen it, and search
3. Share feedback — what worked, what didn't
4. Open an issue — bugs, questions, feature requests
5. Contribute — good first issues listed

See CONTRIBUTING.md | Good first issues: GOOD_FIRST_ISSUES.md

Your AI should remember. Your data should stay yours.

M3 Memory: the foundation for agents that don't forget.