AGM — Agent Memory

Warning — database and Docker: With Docker Compose (cd agent-memory && docker compose up --build), Postgres and Redis are included and the API’s DATABASE_URL is wired in compose — you do not need to set DB URLs in .env for that default run. If you run the API without Docker, you must provide PostgreSQL (pgvector) + Redis and correct DATABASE_URL / REDIS_URL in agent-memory/.env. The live server always needs a real database; skipping that is only possible if you only point MCP or other clients at someone else’s hosted API. Details: docs/DEPLOYMENT.md.

MCP package on PyPI: agent-memory-mcp

Monorepo for agent-memory: a FastAPI service with PostgreSQL (pgvector), Redis, a trust-decay engine, policy rules, optional React dashboard, and an MCP package for tool integrations. Design rationale, problem framing, and related work: docs/RESEARCH.md.

Quick start

git clone https://github.com/OWNER/REPO.git
cd REPO/agent-memory
docker compose up --build

In another terminal (with the API up):

cd agent-memory
PYTHONPATH=. python backend/demo_simulation.py

On Windows PowerShell use $env:PYTHONPATH = "$PWD" instead of PYTHONPATH=..

Architecture

                         +---------------------+
                         |  Browser (:3000)    |
                         |  Vite + nginx       |
                         +----------+----------+
                                    | HTTP
                                    v
+------------+   cache    +---------------------------+
|   Redis    |<-----------|  FastAPI (:8000)          |
| trust/keys |            | agents · memories · stats |
+------------+            | graph · admin · rules     |
      ^                   +-------------+-------------+
      |                                 |
      | async SQLAlchemy                  v
      |                   +---------------------------+
      +-------------------| PostgreSQL + pgvector      |
                          | memories · provenance ·   |
                          | trust_metric_snapshots    |
                          +---------------------------+

Background tasks periodically recompute trust scores and may flag anomalies. See agent-memory/CONTRIBUTING.md for the data-model diagram in more detail.

Documentation

Topic	Location
Docker vs local DB, .env, MCP without hosting DB	docs/DEPLOYMENT.md
Research background (design rationale, related work)	docs/RESEARCH.md
Rules reference & how to add a rule	agent-memory/backend/rules/README.md
Contributing (tests, Docker, model overview)	agent-memory/CONTRIBUTING.md
Service README (ports, API highlights, dev, MCP)	agent-memory/README.md

MCP (Claude Desktop / Claude Code)

The agent-memory-mcp package exposes the HTTP API as MCP tools. Install and configure the stdio server as described in agent-memory/README.md (MCP section) and in agent_memory_mcp/config.py (AGENT_MEMORY_API_URL, AGENT_MEMORY_API_PREFIX).

Example Claude Desktop snippet:

{
  "mcpServers": {
    "agent-memory": {
      "command": "python",
      "args": ["-m", "agent_memory_mcp.server"],
      "env": {
        "AGENT_MEMORY_API_URL": "http://localhost:8000",
        "AGENT_MEMORY_API_PREFIX": ""
      }
    }
  }
}

Replace OWNER/REPO in the Quick start git clone URL with your GitHub org and repository name after publishing.

License: MIT. For the full legal text plus non-binding author requests (citation when you share work publicly, and a quick heads-up for commercial use), see LICENSE.