m3 Memory: memory tool for Hermes Agent

Local-first shared memory for AI agents with hybrid search, MCP tools and a fully offline mode

What m3 Memory does

m3 stores memories in a file on your machine and searches them with FTS5 keyword matching, local BGE-M3 vectors and MMR for diversity. Every tool is a JSON-in, JSON-out CLI call, so MCP is optional and the same store can be scripted from any language, hook or CI job. m3 setup detects installed agents, wires the MCP server and provisions the local embedder, and m3 doctor reports health, memory count and which agents were wired.

A background Cognitive Loop enriches saved memories off the write path, with classification, embeddings and entity extraction, and builds an entity relationship graph. Curation is deterministic: duplicate detection uses cosine similarity against a threshold, decay and pruning use age and signal rules, and applying a curation plan is a single function issuing direct SQL with no model involved. Contradictions are caught on the write path, by a Reflector pass that writes supersedes edges, and by explicit curation plans. It needs no account or API key and can be installed fully offline from pre-staged wheels.

Key features

Hybrid search combining FTS5, local BGE-M3 vectors and MMR
JSON-in, JSON-out CLI tools, with MCP as an option
Contradiction detection on three paths, including supersedes edges
Background Cognitive Loop for enrichment and entity graph building
Shared local embed server reused by every m3 process
Fully offline installation from pre-staged wheels

When to use it

Keeping project decisions available across Claude Code, Gemini CLI and Hermes
Running agent memory on an air-gapped network
Scripting memory writes and searches from CI jobs or hooks

Who it is for: Developers and teams who want one private memory store shared by several agents, including in offline environments.

How it fits with Hermes Agent

m3 is a general memory backend whose project description lists Hermes among the supported agents, rather than a Hermes-only plugin.

How to install m3 Memory

These commands are copied from the project's README. Check the repository for the latest steps before you run them.

pip install m3-memory
m3 setup
m3 doctor

Requirements: Python with pip or pipx; no cloud account, API key or external embedding service is needed

FAQ

What is m3 Memory?

m3 Memory is a local-first memory framework for AI agents. It keeps one shared, searchable store on your machine that tools such as Claude Code, Gemini CLI and other MCP-compatible agents can use.

Does m3 Memory work with Hermes Agent?

The project description lists Hermes among the supported agents. Run m3 setup to detect installed agents and wire the MCP server, then m3 doctor to see which agents were wired.

Is m3 Memory free and open source?

Yes. It is licensed under Apache-2.0 and runs entirely on your machine without an account.