Persistent Memory for AI Agents.
Your agents forget everything between sessions. mem9 fixes that with persistent memory across sessions and machines, shared memory for multi-agent workflows, and hybrid recall with a visual dashboard.

For OpenClaw and ClawHub installs, start here: mem9.ai/openclaw-memory

Server-based memory via mem9-server.

1. Deploy mnemo-server.

   cd server && MNEMO_DSN="user:pass@tcp(host:4000)/mnemos?parseTime=true" go run ./cmd/mnemo-server

2. Install the plugin for your agent (pick one).

   Platform	Install
   Claude Code	/plugin marketplace add mem9-ai/mem9 then /plugin install mem9@mem9
   OpenCode	Add "plugin": ["@mem9/opencode"] to opencode.json
   OpenClaw	Add mnemo to openclaw.json plugins (see openclaw-plugin/README)

3. Provision a tenant and set credentials.

   curl -s -X POST localhost:8080/v1alpha1/mem9s
   # → {"id":"..."}
   
   export MEM9_API_URL="http://localhost:8080"
   export MEM9_API_KEY="..."

   All agents pointing at the same tenant ID share one memory pool.

AI coding agents — Claude Code, OpenCode, OpenClaw, and others — often maintain separate local memory files. As a result:

• 🧠 Amnesia — Agents forget everything when a session ends
• 🏝️ Silos — One agent can't access what another learned yesterday
• 📁 Local files — Memory is tied to a single machine, lost when you switch devices
• 🚫 No team sharing — Your teammate's agent can't benefit from your agent's discoveries

mnemos gives every agent a shared, cloud-persistent memory with hybrid vector + keyword search — powered by TiDB Cloud Starter.

mnemos uses TiDB Cloud Starter (formerly TiDB Cloud Serverless) as the backing store for mnemo-server:

This architecture keeps agent plugins stateless — all state lives in mnemo-server, backed by TiDB.

mnemos provides native plugins for major AI coding agent platforms:

All plugins expose the same 5 tools: memory_store, memory_search, memory_get, memory_update, memory_delete.

A key design principle: agent plugins carry zero state. All memory lives in mnemo-server, backed by TiDB/MySQL. This means:

• Agent plugins stay stateless — deploy any number of agent instances freely; they all share the same memory pool via mnemo-server
• Switch machines freely — your agent's memory follows you, not your laptop
• Multi-agent collaboration — Claude Code, OpenCode, OpenClaw, and any HTTP client share memories when pointed at the same server
• Centralized control — rate limits and audit live in one place

The broader mem9 product surface is split across a few repositories and workstreams:

Notes:

• dashboard/app/ in this repo is the frontend half of the dashboard product. The backend half for async analysis lives in mem9-node, especially its apps/api and apps/worker.
• mem9-tester is not part of the production runtime. It exists to validate installation / reconnect behavior and reduce fragile manual testing.
• If someone says mem9-benchmark, they usually mean the benchmark work around mem9 evaluation. In this repo, that work currently starts in benchmark/.

Agent identity: X-Mnemo-Agent-Id header.

Minimal runtime config is MNEMO_DSN. Everything else is optional or only applies to specific deployment modes.

These are only relevant when MNEMO_ENCRYPT_TYPE=kms. The server uses the AWS SDK default config chain; the common environment-based inputs referenced in code are:

make build
cd server
MNEMO_DSN="user:pass@tcp(host:4000)/mnemos?parseTime=true" ./bin/mnemo-server

docker build -t mnemo-server ./server
docker run -e MNEMO_DSN="..." -p 8080:8080 mnemo-server

Vector Clock CRDT was deferred and removed from the roadmap.

See CONTRIBUTING.md for development setup and guidelines.

Apache-2.0

_{Built with TiDB Starter — zero-ops database with native vector search.}