Home

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 layers in personas focused on a specific kind of work — activation shipped, so you can invoke one explicitly or let a workflow step select it.

AgentM works best paired with crickets — the toolkit of plugins (capabilities, skills, hooks, and sub-agents) that make it even more useful.

🚀 Get started

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_ROOT="<your-google-drive>/agent"

# Install AgentM for every project on this machine
bash ~/agentm/install.sh

# Add the crickets plugins (Claude Code + Antigravity)
curl -fsSL https://raw.githubusercontent.com/alexherrero/crickets/main/bootstrap.sh | bash

Step-by-step, with the vault explained: Install AgentM. More on the available configurations here.

📖 Learn more

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.
• What we built, when — the combined build timeline for AgentM and crickets, plus the roadmap-era-to-release-tag decoder.

🗂️ Browse by type

The memory vault itself is browsable by type. Every memory appears on a generated map beside the others of its type (preference, convention, fix, and the three more the filing contract registers), so you can scan a type without a search, and moc-root.md is the map to start from. See Memory daemon reference for how the maps are generated.

🤖 Sub-agents

A sub-agent is a helper your coding agent hands a narrow job to. It gathers what you asked for and reports back without touching your code.

AgentM ships two sub-agents, both serving the memory engine:

Sub-agent	What it does
adapt-evaluator	Judges a candidate skill against the adapt-don't-import rule and writes the watchlist entry you review.
memory-idea-researcher	Fills out an idea-incubator entry, reading the vault, your notes, and the web under set budgets.

Both install into your agent's .claude/agents/ folder — or .agents/skills/ on Antigravity. Alongside them, AgentM installs its personas, compiling named stances like the engineer or the reviewer into that same directory.

Any other sub-agent you see named — such as explorer, adversarial-reviewer, or documenter — comes from crickets, not AgentM. Standalone AgentM includes only the two memory sub-agents and the personas. The review, phase-loop, and wiki agents arrive with crickets plugins. See Toolkit interface for where the line sits.

---

Note

Latest release: v10.8.0. Memory keeps only the notes that earned their place: the memory junk review's 121 purged on your typed count, 22 moved with their links repaired, and 6 duplicate lessons folded into their keepers, taking curated memory from 342 notes to 205 with every lesson's sources still resolving. The writers that made the junk are fixed: an issue reference takes the repository its text names, a page's slices are one source for a lesson, and a superseded lesson retires to the archive once a live successor answers instead. Four adversarial reviews hardened the fixes before the release.