One shared memory and coordination layer for every AI coding tool you run.
You run Claude Code, Codex, Gemini CLI, opencode, Antigravity, MiMo Code — and every one of them starts every session knowing nothing about what the others just did.
AMH fixes that. One local directory. One ledger. Every agent reads and writes the same memory, messages, and tasks.
You asked Codex to refactor auth/. Then you opened Claude Code to write tests. It has no idea:
- what Codex changed,
- which decision you rejected and why,
- that there's a blocked task waiting on a review,
- that Gemini already tried that approach and it failed.
So you re-explain. Every. Single. Time.
Not another vector database. Not another MCP memory server you have to trust and babysit.
AMH is a local-first ledger + coordination bus that sits under your AI tools.
| Shared memory | Agents write what they learned. Anyone can search it. Core / working / archive tiers, FTS5 full-text (works for CJK too). |
| Agent radio | Cross-tool messaging with replyTo threading. One agent can leave a note for another. |
| Shared task queue | open → claimed → in_progress → blocked → done, with handoff notes and quality gates. |
| Workflows | Multi-role pipelines (planner / executor / reviewer / observer) with node-level history. |
| Roles & teams | 6 job roles, 7 registered agents, member-of relations, live in the dashboard. |
| Approval gates | Machine-readable policy checks that run before dispatch, not after you've been burned. |
| Dispatch | Hand pending work to a verified CLI runner, optionally in an isolated git worktree. |
| Dashboard | Local React UI for all of the above. |
- ❌ No LLM proxying — each tool talks to its own API with its own key
- ❌ No unified config — your tools stay independently configured
- ❌ No token scraping — AMH does not read other tools' API keys. A credential profile is read only when you store one and a dispatch asks for it.
- ❌ No cloud — everything lives in
~/.ai-memory
That last row is the whole point for a lot of people.
⚠️ Do not install theai-memory-hubpackage from npm. That name is registered to an unrelated project. This project is not published to npm yet. Install it from source:
git clone https://github.com/monkey-sking/ai-memory-hub.git
cd ai-memory-hub
npm install && npm link
amh initRequires Node 24+.
# 1. See what's already there
amh status
# 2. Write something you don't want to re-explain
amh record "Chose Postgres over SQLite for the events table because of concurrent writers" \
--source claude --kind decision --project myapp --tags db,architecture
# 3. Make it searchable (required after every write)
amh sync && amh index
# 4. Prove another agent can find it
amh search "why postgres"
# 5. Leave a message for a different tool
amh radio send --from claude --to codex "Don't touch auth/ until the migration lands"
# 6. Open the dashboard
amh appPoint any of your AI tools at it — they all speak the same CLI:
amh search "<what am I working on>" # read before you act
amh record "<what you learned>" --source <your-name>
amh task list # is anything assigned to meThis isn't a demo repo. It's the thing one developer actually runs every day:
- 813 memory events
- 52 cross-agent radio messages
- 66 tasks (46 done, 20 active)
- 7 registered agents:
codex,claude,gemini,antigravity,opencode,mimocode,workbuddy
Verified via amh status on the maintainer's machine.
| Supermemory | Mem0 | Basic Memory | AMH | |
|---|---|---|---|---|
| Storage | Cloud (self-host on Scale) | Cloud only | Local markdown | Local SQLite + JSONL |
| Needs an account / API key | Yes | Yes | No | No |
| Touches your LLM traffic | Yes | Yes | No | No |
| Memory search | ✅ | ✅ | ✅ | ✅ |
| Inter-agent messaging | ❌ | ❌ | ❌ | ✅ |
| Shared task queue | ❌ | ❌ | ❌ | ✅ |
| Workflows & roles | ❌ | ❌ | ❌ | ✅ |
| Approval gates | ❌ | ❌ | ❌ | ✅ |
| Works without MCP | ❌ | ❌ | ❌ | ✅ (plain CLI) |
The one-liner: they sell remembering. AMH sells stopping five agents from working blind.
Claude Code · Codex · Gemini CLI · Antigravity (agy) · opencode · MiMo Code · Marvis · QClaw · OpenClaw · WorkBuddy · and anything that can shell out.
- Full documentation:
docs/ - Chinese readme:
README.zh-CN.md - Contributing:
CONTRIBUTING.md
Apache-2.0