Home

The agent harness that gives you the assistant you want — part Star Trek Computer, part J.A.R.V.I.S.

_{Works with Claude Code + Antigravity — see compatibility}

Agent M is a phase-gated agent harness with a persistent memory layer. The harness gives the dev loop hard boundaries — Setup · Plan · Work · Review · Release · Bugfix — so an agent executes one phase at a time against on-disk state instead of freestyling the whole lifecycle. The MemoryVault gives it a durable, file-based memory that carries your preferences, project state, and learned lessons across sessions and across projects. Imagine the workflows you saw in the movies: you talk to your agent, it remembers your projects and your notes, and it improves automatically as you work — no knowledge graph to hand-maintain. Agent M has grown across the paired releases of agentm (this harness) and crickets (the toolkit of skills, hooks, and sub-agents that ride on top).

Note

This wiki documents the agentm repo for contributors. Projects that install the harness get templates/wiki/ scaffolded into them instead — see ADR 0002 for why the two are kept separate.

📚 Get started

Agent M is two sibling repos plus a vault folder. Clone both, point the vault at your sync setup, and the harness is operational.

• Tutorial — your first harness install — fresh clone to a healthy installed scratch project in ~5 minutes.
• Install the harness into a project — add the scaffold to an existing repo.
• Run without a vault — operate the harness with no MemoryVault configured (repo-local state).

🔧 What do you want to do?

What	How-to
🧱 Stand up a project — detect → propose → approve → persist the per-project config	Configure a new project
⬆️ Pull a newer harness into a project that already has one	Update an installed harness
🚀 Cut a release — tag, changelog, GitHub release	Cut a release
🧠 Tune the memory — recall budgets, save modes, confidence thresholds per phase	Use auto-context in phases · Tune auto-orchestration
🩺 Keep the vault healthy — lint it, find missing note links	Audit the vault · Find missing note links
🌐 Read the vault from any agent — Claude.ai · Gemini · ChatGPT · Antigravity	Use AgentMemory in any agent
🖥️ Keep state per-repo — when to stay --scope project	Use per-project install
🔌 Expose the vault as an MCP server — Claude Code · Cursor · Goose · Claude Desktop	Stand up the memory MCP server

📖 Look up a detail

For contributors running the harness:

• Compatibility — supported hosts (Claude Code · Antigravity) + OS matrix + the adapter contract.
• Installer CLI — flags, prerequisites, and the ownership table for install.sh / install.ps1.
• CI gates — what each CI workflow proves and the script behind it.
• Completed features — the reverse-chronological log of shipped work.
• Memory MCP tools — four-tool MCP surface: parameters, pagination, soft-delete contract, error codes.

Note

Latest release — v5.5.0 (2026-06-18): V5-3 storage cutover — vault backend deleted from the harness_memory.py kernel; harness_state_dir / read_state_file / write_state_file always device-local; phase_recall always ""; resolve_documenter_context always None; vault_path() raises StorageBackendNotInstalledError when storage.backend=vault + no vault; personal-private/ → personal/ group rename in lockstep. (ADR 0018)

→ Full field-level detail lives under Reference in the sidebar.

🏛️ How it's built

The structural component map — six components, each a folder under Architecture in the sidebar. → Browse the architecture

• AgentMemory — the MemoryVault substrate: durable, file-based memory across sessions and projects.
• Device-Wide Substrate — how the harness and its memory span every project on the machine.
• Phases — the phase-gated workflow with hard boundaries and on-disk state.
• Orchestration and Auto-Detection — auto-wiring the right phase + context by detecting a project's shape.
• Host adapters — adapting to each host (Claude Code · Antigravity 2.0 · Antigravity CLI).
• Toolkit interface ↔ crickets — the seam with the sibling toolkit: what each owns, how they compose.

🧩 Major designs

The high-level designs behind Agent M's memory layer — the full HLDs, where the design started and where it's going. → Browse all designs

• MemoryVault — permanent agent memory: write-primitives · recall-loop · reflection-and-recovery · idea-ledger · seed-pass · discovery-mining.
• Agent Memory Evolution V1→V4 — the full HLD: where Agent M started, how it grew, where it's going.
• Device-Wide Architecture — the device-wide design Agent M targets.
• Memory-OS Architecture (V5) — the V5 memory-OS HLD.

💡 Why it works the way it does

• Product intent — what problem the harness solves and for whom.
• Auto-detect + auto-configure — why first-session config proposes-then-approves and lives in project.json.
• How the pieces fit — the narrative of how phases, adapters, templates, and scripts interact.
• GitHub Projects integration — why and how the harness writes to ProjectsV2.
• Auto-orchestration — why the memory skills became a push surface that never nags.
• Single-repo state mode — how the harness degrades to repo-local state when no vault is reachable.
• Memory↔process seam — why a process reads memory through a one-way, read-only client instead of reaching into the engine.

📐 Architecture decisions

Every load-bearing call is recorded as an ADR — the "why X, and why not Y" trail, with re-audit triggers. → Browse all decisions