# m3-memory

> Local-first Agentic Memory Layer for MCP Agents • 25 tools • Hybrid search (FTS5 + vector + MMR) • GDPR • 100% local

- **URL**: https://www.freshcrate.ai/projects/m3-memory
- **Author**: skynetcmd
- **Category**: MCP Servers
- **Latest version**: `v2026.6.1.0` (2026-06-01)
- **License**: MIT
- **Source**: https://github.com/skynetcmd/m3-memory
- **Language**: Python
- **GitHub**: 10 stars, 1 forks
- **Registry**: github
- **Tags**: `agentic-memory`, `ai-agents`, `aider`, `claude-code`, `gdpr`, `gemini-cli`, `hybrid-search`, `local-llm`, `mcp`, `python`

## Description

Local-first Agentic Memory Layer for MCP Agents • 25 tools • Hybrid search (FTS5 + vector + MMR) • GDPR • 100% local

## Recent releases

| Version | Date | Urgency | Changes |
| --- | --- | --- | --- |
| `v2026.6.1.0` | 2026-06-01 | High | Faster, more resilient, more efficient and more robust while maintaining backward compatibility. FIPS 140-3 ready. Local first, cloud capable memory engine. |
| `v2026.5.30.0` | 2026-05-30 | High | ## Entity coalescing v2 (reversible apply/unapply) + catalog registration + CLI exit-code fix  ### Added  - **Entity-coalescing v2 — reversible overlay apply/unapply** (`files_entity_coalesce_apply` / `files_entity_coalesce_unapply` MCP tools; `entity-coalesce-apply` / `entity-coalesce-unapply` CLI). Materializes reviewed/auto-merge candidates as a reversible `same_as` + shared-`cluster_id` overlay — members are never deleted, the canonical view is a read-time projection, a deterministic represe |
| `v2026.5.21.0` | 2026-05-21 | High | ## v2026.5.21.0 — green-CI maintenance release  First release on fully-green CI (Lint, Type Check, Security, and Tests on ubuntu/macos/windows all passing).  ### Fixes - **F821 undefined-name errors (30)** — the Phase 7+8 modularization left dangling symbol references in `memory/write.py`, `search.py`, `enrich.py`; restored the missing imports and re-exports. - **Mypy type errors (64)** — resolved across `bin/`: missing `memory_core` re-exports, `m3_core_rs` typing, exception-base annotations, h |
| `v2026.5.18.1` | 2026-05-18 | High | **Full Changelog**: https://github.com/skynetcmd/m3-memory/compare/v2026.5.18.0...v2026.5.18.1 |
| `v2026.5.6.3` | 2026-05-07 | High | ## Highlights  - **GitHub Actions Node-24 upgrade.** Pins every workflow action to its latest Node-24-compatible release-tag SHA so CI keeps working past GitHub's Node 20 removal on 2026-09-16.  \| action \| from \| to \| \| --- \| --- \| --- \| \| `actions/checkout` \| v4 \| **v6.0.2** \| \| `actions/setup-python` \| v5 \| **v6.2.0** \| \| `actions/upload-artifact` \| v4 \| **v7.0.1** \| \| `actions/download-artifact` \| v4 \| **v8.0.1** \|  - **New banner art** at `docs/M3-banner.jpg`.  ## Migration notes  Major-vers |
| `v2026.5.1.1` | 2026-05-01 | High | ## Highlights  This release rolls up roughly five weeks of ingest-pipeline work plus a documentation, hygiene, and security pass.  - **Tool count: 66 → 72.** Six new MCP tools: `entity_search`, `entity_get`, `extract_pending`, `enrich_pending`, `memory_search_routed`, `memory_write_from_file`. See [`docs/MCP_TOOLS.md`](https://github.com/skynetcmd/m3-memory/blob/main/docs/MCP_TOOLS.md). - **`bin/m3_enrich.py` CLI** — first-class user-facing enrichment over the Observer/Reflector pipeline. Durabl |
| `v2026.4.24.12` | 2026-04-25 | High | ## Two pain points caught testing the plugin on Windows  ### Plugin commands now self-resolve  Caught when `/m3:doctor` ran on Windows: `pip install --user m3-memory` puts `mcp-memory.exe` at `%APPDATA%\Python\Python<NN>\Scripts`, which is **not** on Windows PATH by default. The plugin command had Claude spelunking through 8 fallback paths before finding `python -m m3_memory.cli` works.  `/m3:doctor`, `/m3:status`, and `/m3:install` now explicitly list the resolver chain in their command bodies: |
| `v2026.4.20` | 2026-04-17 | High | ## Fixes - Resolve SQLite deadlock in `memory_consolidate` via connection reuse in `memory_link_impl` - Resolve database contention and sync hangs on large DBs (migration timeout bump, SAVEPOINT-wrapped pg_sync, VACUUM skip for >500MB DBs, WAL contention fix in `record_history`)  ## Features - `memory_search` gains `recency_bias` (float) and `adaptive_k` (bool) params - `conversation_search` auto-pairs user turns with adjacent assistant replies (0.85x score) - `precedes` / `follows` added to `VA |
| `v2026.4.19` | 2026-04-16 | High | Fix PyPI rendering: absolute URLs for all links and images, markdown banner fallback, incorporate doc edits. |
| `v2026.4.18` | 2026-04-16 | High | Fix PyPI page rendering: banner image, absolute links, demo SVGs. |

## Citation

- HTML: https://www.freshcrate.ai/projects/m3-memory
- Markdown: https://www.freshcrate.ai/projects/m3-memory.md
- Dependencies JSON: https://www.freshcrate.ai/api/projects/m3-memory/deps

_Generated by freshcrate.ai. Indexes github releases for AI-agent ecosystem packages._