The memory that follows your coding agent — so you get the assistant you actually wanted.
Works with Claude Code + Antigravity — see compatibility
AgentM gives your coding agent a permanent memory. It writes what it learns — about you, and about each project — as plain Markdown notes in a folder you own: no hidden database, nothing you can't read or edit yourself. Memory splits into two tenants — personal notes on how you like to work, and a per-project record of decisions and open threads — so a fresh session can pick up exactly where the last one left off. It brings the right notes back at the right moment, and looks after the collection over time so it gets better instead of messier.
Said differently, AgentM is the memory and self-improvement layer — it dreams and learns between sessions. A phase-gated development workflow (setup · plan · work · review · release · bugfix) can run on top of it, provided by a companion toolkit rather than by AgentM itself. Its customization system extends AgentM's own abilities through plugins that enable long-running, nuanced development workflows, automated project management, and more. Its opinion system is designed to layer in personas focused on a specific kind of work — activation is the next build slice.
AgentM works best paired with crickets — the toolkit of plugins (capabilities, skills, hooks, and sub-agents) that make it even more useful.
AgentM installs alongside your coding agent — Claude Code or Antigravity — so you can be up and running in a few minutes. See requirements.
Install AgentM and crickets with the recommended configuration and a Google Drive vault:
# Point the vault at a Google Drive folder named "Agent" (the recommended default)
export MEMORY_VAULT_PATH="<your-google-drive>/Agent"
# Install AgentM for every project on this machine (user scope) + verification hooks
bash ~/agentm/install.sh --hooks --scope user
# Add the crickets plugins (Claude Code + Antigravity)
curl -fsSL https://raw.githubusercontent.com/alexherrero/crickets/main/bootstrap.sh | bashMore on the available configurations here.
The wiki covers everything there is to know about AgentM. A few links to get you started.
- Why we built it — the problem, the solution and the reason.
- Architecture — how it's made.
- Reference — fields, flags, schemas and more.
Note
Latest release: v9.5.0. A recalled memory's score breakdown used to exist for one prompt turn, then vanish — memory-recall trace <slug> now explains why any past recall surfaced what it did, kept in the existing per-recall ledger.
Contributions are welcome — see CONTRIBUTING.md for information on how to contribute to AgentM.
Multi-licensed so each layer carries the license that fits it: code under Apache-2.0, content — docs, prompts, prose, and other .md — under CC-BY-4.0, and the "AgentM" name and brand under trademark. Prompt text embedded in a code file counts as content, even though it lives there. Both licenses allow commercial use and derivative works with attribution — see NOTICE for the attribution string.