m3-memory 2026.4.6

Local-first agentic memory layer for MCP agents — 25 tools, hybrid search, GDPR, no cloud

🧠 M3 Memory — Local-First Agentic Memory for MCP Agents

Give your MCP agents persistent, private memory that actually works. No cloud. No forgetting.

Industrial-strength memory layer that just works. M3 Memory gives AI agents persistent, private, intelligent memory — running 100% on your hardware, with no cloud APIs and no subscriptions.

• 🛠️ 25 MCP tools — write, search, link, graph, verify, sync, export, and more
• 🔍 Hybrid search — FTS5 keywords + vector similarity + MMR diversity re-ranking
• 🚫 Contradiction detection — stale facts are superseded automatically
• 🔄 Cross-device sync — SQLite ↔ PostgreSQL ↔ ChromaDB, bi-directional
• 🛡️ GDPR-ready — Article 17 (forget) and Article 20 (export) built in
• 🔒 Fully local — your embeddings, your hardware, your data

Works with Claude Code, Gemini CLI, Aider, OpenClaw, or any MCP-compatible agent.

⭐ Star if you want local agents that remember — feedback & issues very welcome!

---

Table of Contents

• Why M3 Memory
• Architecture
• Quick Start
• Features
• 25 MCP Tools
• Documentation
• Contributing

---

Why M3 Memory

Most agent memory solutions require you to pick one: local speed, or cloud persistence, or multi-agent sharing. M3 Memory gives you all three, with zero data leaving your machine by default.

Example: You're debugging a deployment issue on your MacBook at a coffee shop. Claude Code recalls the architecture decisions from last week, the server configs from yesterday, and the troubleshooting steps that worked before — all from local SQLite, no internet required. Later, at your Windows desktop at home, Gemini CLI picks up exactly where you left off. Same memories, same context, same knowledge graph — synced in the background the moment your laptop hit the local network.

⭐ Your AI's memory belongs to you, lives on your hardware, and follows you across every device and every agent.

---

Architecture

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

The Memory Write Pipeline

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>

---

Quick Start

Prerequisites

• Python 3.11+
• Any OpenAI-compatible local LLM server: LM Studio, Ollama, vLLM, LocalAI, llama.cpp
• (Optional) PostgreSQL + ChromaDB for full cross-device federation

Install

Option A — pip (quickest):

pip install m3-memory

Option B — clone (for development or full feature set):

git clone https://github.com/skynetcmd/m3-memory.git
cd m3-memory

python -m venv .venv
source .venv/bin/activate          # macOS/Linux
# .\.venv\Scripts\Activate.ps1    # Windows PowerShell

pip install -r requirements.txt

Validate & Test

python validate_env.py             # Check all dependencies and LLM connectivity
python run_tests.py                # Run the end-to-end test suite

Connect Your Agent

Copy mcp.json.example to your agent's MCP config location and update the cwd path:

{
  "mcpServers": {
    "memory": {
      "command": "python",
      "args": ["bin/memory_bridge.py"],
      "cwd": "/path/to/m3-memory"
    }
  }
}

Agent	Config location
Claude Code	~/.claude/claude_desktop_config.json or .mcp.json in project root
Gemini CLI	~/.gemini/settings.json
Aider	.aider.conf.yml (via --mcp-server flag)

For OS-specific setup: macOS | Linux | Windows

M3 Memory can also be discovered automatically in Claude Code and other MCP clients via the MCP Registry. See mcp-server.json for the registry manifest.

---

Features

🔍 Hybrid Search That Actually Works

Three-stage pipeline consistently outperforms pure vector search:

1. FTS5 keyword — BM25-ranked full-text with injection-safe sanitization
2. Semantic vector — cosine similarity on 1024-dim embeddings via numpy
3. MMR re-ranking — Maximal Marginal Relevance ensures diverse results — no more five near-identical memories

Every result can return a full score breakdown (vector, BM25, MMR penalty) via memory_suggest.

🚫 Contradiction Detection

Write a fact that conflicts with an existing one — M3 detects it automatically. The old memory is soft-deleted, a supersedes relationship is recorded, and the full history is preserved. No stale data, no manual cleanup.

🕸️ Knowledge Graph

Memories form a web. M3 auto-links related memories on write (cosine > 0.7) and supports 7 relationship types: related, supports, contradicts, extends, supersedes, references, consolidates. Traverse up to 3 hops with a single memory_graph call.

🧹 Self-Maintaining

• Importance decay — memories fade 0.5%/day after 7 days unless reinforced
• Auto-archival — low-importance items (< 0.05) older than 30 days move to cold storage
• Per-agent retention — set max memory count and TTL per agent
• Consolidation — local LLM merges old memories into summaries when groups grow large
• Deduplication — configurable cosine threshold catches near-duplicates

⏳ Bitemporal History

Track not just when a fact was stored, but when it was actually true. Query with as_of="2026-01-15" to see the world as your agent knew it on that date — essential for compliance and debugging.

🤖 LLM-Powered Intelligence (Local)

Any OpenAI-compatible server works (LM Studio, Ollama, vLLM, LocalAI):

• Auto-classification — pass type="auto" and the LLM categorizes into 18 types
• Conversation summarization — compress long threads into 3-5 key points
• Multi-layered consolidation — merge related memory groups into summaries

Zero API costs. Zero data exfiltration.

🛡️ Security & Compliance

Layer	Protection
Credentials	AES-256 encrypted vault (PBKDF2, 600K iterations). OS keyring integration. Zero plaintext storage.
Content	SHA-256 signing on every write. memory_verify detects post-write tampering.
Input	Poisoning prevention rejects XSS, SQL injection, Python injection, and prompt injection at write boundary.
Search	FTS5 operator sanitization prevents query injection.
Network	Circuit breaker (3-failure threshold). Strict timeouts. Tokens never logged.

GDPR-Ready:

• gdpr_forget — hard-deletes all data for a user (Article 17)
• gdpr_export — returns all memories as portable JSON (Article 20)

🔄 Cross-Device Sync

• Bi-directional delta sync: SQLite ↔ PostgreSQL via UUID-based UPSERT
• Crash-resistant — watermark-based tracking, at-least-once delivery
• ChromaDB federation for distributed vector search across LAN
• Hourly automated sync; manual sync via chroma_sync tool

---

25 MCP Tools

Category	Tools
Memory Ops	memory_write, memory_search, memory_suggest, memory_get, memory_update, memory_delete, memory_verify
Knowledge Graph	memory_link, memory_graph, memory_history
Conversations	conversation_start, conversation_append, conversation_search, conversation_summarize
Lifecycle	memory_maintenance, memory_dedup, memory_consolidate, memory_set_retention, memory_feedback
Data Governance	gdpr_export, gdpr_forget, memory_export, memory_import
Operations	memory_cost_report, chroma_sync

---

Documentation

File	Purpose
CORE_FEATURES.md	Feature overview — start here
ARCHITECTURE.md	Agent instruction manual: all 25 MCP tools, protocols, usage rules
TECHNICAL_DETAILS.md	Deep-dive: storage internals, search pipeline, schema, sync, security
ENVIRONMENT_VARIABLES.md	Security configuration and credential setup
CHANGELOG.md	Release history and what changed
CONTRIBUTING.md	How to contribute, run tests, and fix issues

---

Community

Join the M3-Memory Discord server to get help, share your setup, discuss ideas, and follow development.

Channel	Purpose
#start-here	New? Start here — overview & quick links
#ask-anything	Setup help, config questions, how-tos
#bug-reports	Report issues with steps to reproduce
#showcase	Share your M3-Memory setups and demos
#search-quality	Hybrid search tuning & benchmarks
#sync-federation	Multi-device sync & ChromaDB federation
#memory-design	Architecture discussions & research

M3_Bot is live in the server — mention @M3_Bot or use !ask <question> in any channel to get answers straight from the documentation.

---

Contributing

See CONTRIBUTING.md for how to get started, run the test suite, and submit changes.

---

Project Structure

bin/          Core MCP bridges, SDK, and automation scripts
memory/       SQLite database and migration logic
config/       Configuration templates for agents and shell
docs/         Architecture diagrams, API reference, and OS install guides
scripts/      Maintenance and utility scripts (fix_bugs, fix_db, fix_lint, etc.)
logs/         Centralized audit and debug logs
tests/        End-to-end test suite

---

Status: Production Release — v2026.04 · MIT License · Changelog

---

---

M3 Memory: the industrial-strength foundation for agents that remember.