agent-memory

Project Memory

This repository keeps canonical project memory in .agent-memory/state.json.

History and checkpoints live in .agent-memory/history/.

Readable projections live in docs/agent-memory/.

Recommended reading order:

1. docs/agent-memory/README.md
2. docs/agent-memory/project-map.md
3. docs/agent-memory/current-focus.md
4. docs/agent-memory/gotchas.md when debugging gets noisy or surprising
5. docs/agent-memory/next-steps.md when you need a clean starting point

Use npx agent-memory sync, npx agent-memory recall, and npx agent-memory query to maintain and retrieve project memory.

Durable project memory with history, recall, and query.

agent-memory is now a full memory system for repositories:

• it stores the current canonical memory in /.agent-memory/state.json
• it records durable history in /.agent-memory/history/
• it supports configurable recall policies in /.agent-memory/config.json
• it projects the current memory into docs/agent-memory/
• it lets you consolidate memory with recall
• it lets you retrieve memory with query
• it lets you inspect backlog and checkpoint drift with status
• it can automate import-sync and recall with automate
• it can integrate Claude Code and Codex with integrate

Why This Exists

Repositories keep losing the same expensive context:

• architectural boundaries live across code, manifests, and half-finished docs
• current state gets trapped in chats and PR threads
• repeated gotchas are rediscovered instead of remembered
• long-lived projects need both memory maintenance and memory retrieval

agent-memory turns that into a structured repository memory system instead of a pile of static notes.

Core Model

The system now has four persistent layers:

• /.agent-memory/state.json Current canonical memory bundle
• /.agent-memory/history/events.jsonl Append-only history of tool runs and imported sessions
• /.agent-memory/history/checkpoints/ Bundle checkpoints written after init, update, and recall
• /.agent-memory/sources.json Registered external history sources
• /.agent-memory/config.json Recall defaults, policy, and backlog thresholds

Readable projections still live in:

• docs/agent-memory/README.md
• docs/agent-memory/project-map.md
• docs/agent-memory/current-focus.md
• docs/agent-memory/gotchas.md
• docs/agent-memory/next-steps.md

An entry block is also written into the preferred top-level entry file.

Installation

Install agent-memory from the npm registry as a project dependency:

npm install -D @agent-connect/memory

or:

pnpm add -D @agent-connect/memory

After installation, use the local CLI with npx agent-memory ....

How It Works

agent-memory no longer treats markdown files as the source of truth.

Instead, it:

1. collects repository context
2. builds or refreshes a canonical bundle
3. appends durable history events and checkpoints
4. projects the active bundle into repository docs
5. lets you consolidate history back into memory with recall
6. lets you ask memory questions with query
7. lets you inspect backlog and checkpoint drift with status
8. can run local automation for import-sync and recall with automate
9. can integrate chat clients and MCP tooling with integrate

If you need to control the runtime used for synthesis, use --provider=auto|codex|claude.

Commands

Bootstrap memory

npx agent-memory init

Creates a fresh canonical state, resets the history scaffold, writes the first checkpoint, and projects the bundle into docs.

Refresh current memory

npx agent-memory update

Refreshes the active canonical bundle from current repository evidence and writes a new checkpoint plus tool-run event.

Consolidate memory

npx agent-memory recall

Reads unrecalled history, proposes a consolidated bundle, shows summary changes and file diffs, and applies only after confirmation. The preview also includes a grouped unrecalled history summary before apply.

If no unrecalled events produced durable changes, recall exits with a clear no-op message and does not write a checkpoint or tool-run event.

• --section=... to limit consolidation to part of the bundle
• --policy=... to apply policy presets such as imports-only or project-map protection
• --show-diff to expand from summary-first preview into file-level diffs

Query memory

npx agent-memory query "how does caching work?"

Returns a short answer plus citations from bundle sections, history events, and checkpoints. Natural-language structured questions such as what changed recently?, what should I do next?, and what are the known traps? are supported through the same command.

If current memory cannot support a confident answer, query now returns an explicit evidence-insufficient response instead of bluffing.

For agent-facing automation you can request JSON output:

npx agent-memory query "what changed recently?" --output=json

Bundle citations can also be cross-linked to the generated projection docs under docs/agent-memory/. Projects can override retrieval instructions per query mode through .agent-memory/config.json.

Add external session sources

npx agent-memory add claude-local ~/.claude --name claude

Registers an external session source.

Sync external session sources

npx agent-memory sync --all

Normalizes external sessions into durable history events.

sync may partially succeed:

• imported sessions become history events
• duplicate sessions are skipped
• broken session files are reported as failures without aborting the whole source

Each source also records sync status, last imported count, and the last sync error when applicable.

Inspect memory status

npx agent-memory status

Shows:

• state and latest checkpoint id
• unrecalled backlog counts plus a grouped summary of unrecalled history
• source sync health
• checkpoint drift summary
• retention and archive summary
• the next suggested action

Local automation daemon

npx agent-memory automate start
npx agent-memory automate status
npx agent-memory automate run-once
npx agent-memory automate stop

The first Phase 4 milestone is a local built-in automation daemon.

• it runs as a repo-local background process
• it can run sync --all and recall --yes on a schedule
• it writes runtime state under .agent-memory/automation/
• it records the latest machine-readable run result in .agent-memory/automation/latest-run.json
• it applies retention pruning by default and archives aged history/checkpoints under .agent-memory/archive/
• it uses aggressive auto-apply recall by default
• dirty worktrees do not block automation cycles in this first version

Retention behavior in this first policy-controls milestone:

• retention is enabled by default
• pruning only happens inside automation cycles
• pruning is archive-first, not direct hard delete
• archived batches later expire automatically from .agent-memory/archive/
• archived data does not participate in active query, recall, or status baselines

Integrate Claude Code + Codex

npx agent-memory integrate
npx agent-memory integrate --dry-run
npx agent-memory integrate --status
npx agent-memory integrate --status --output=json
npx agent-memory integrate --repair

This command is separate from init.

• init stays repo-local and never edits global chat tooling config
• integrate writes Claude Code project files
• integrate also registers Codex MCP globally with a safe merge

Install UX notes:

• --dry-run previews project/user changes without writing files
• --status is read-only and summarizes Claude/Codex integration health
• --status --output=json returns machine-readable integration status
• --repair only repairs managed mismatches; it does not create missing components
• --dry-run and --status never write project files or global Codex config
• normal integrate may write user-scope Codex config when Codex integration is selected

Generated/updated files include:

• .mcp.json
• .claude/settings.json
• .claude/skills/agent-memory/SKILL.md
• AGENTS.md
• ~/.codex/config.toml

Claude Code integration uses project MCP + project skills + SessionStart and Stop hooks. Codex integration uses MCP + AGENTS.md + the local daemon.

The generated Claude skill and Codex guidance now prefer the higher-level MCP workflow tools:

• memory_assess for a quick health and readiness check
• memory_compact_handoff before compact or session handoff
• memory_maintain for a one-shot maintenance pass
• lower-level tools such as memory_query, memory_status, memory_validate, and automation_* only when finer control is needed

MCP server

npx agent-memory mcp
npx agent-memory mcp --transport=http --port=8765

This starts the local MCP server used by Claude Code and Codex integrations.

• stdio remains the default and is what integrate writes into .mcp.json
• optional HTTP transport is available with --transport=http --port=<port>
• HTTP mode also supports --host and --allowed-hosts

The MCP server exposes:

• high-level workflow tools: memory_assess, memory_compact_handoff, memory_maintain
• lower-level retrieval and health tools: memory_query, memory_status, memory_validate
• automation controls: automation_status, automation_ensure_running, automation_run_once

Tool calls now return human-readable text content plus versioned structuredContent envelopes for MCP clients, and long-running tools emit progress notifications.

Ensure automation is running

npx agent-memory automate ensure-running

This is the lightweight startup command used by Claude Code SessionStart hooks and can also be called directly.

Self-host dogfood

This repository can also exercise itself as a long-lived dogfood arena.

npm run dogfood:init
npm run dogfood:exercise
npm run dogfood:repair
npm run dogfood:status

Optional provider override:

AGENT_MEMORY_DOGFOOD_PROVIDER=claude npm run dogfood:init

Current dogfood behavior:

• repo-root .agent-memory/, docs/agent-memory/, .mcp.json, .claude/, and AGENTS.md form the stable self-host baseline
• dogfood:init installs the published @agent-connect/memory package into the temporary repo clone before using npx agent-memory ...
• exercise runs in an isolated git worktree under temp/dogfood/worktree
• dogfood now inherits your real HOME by default, so Codex and Claude use the same user-level auth/config they would during normal local work
• tests and other isolated runs can still override HOME explicitly when they need a sandbox
• dogfood structured runs now do a real provider preflight instead of trusting --version alone, so auto can fall back when one backend is installed but not authenticated
• reports are written under temp/dogfood/reports/
• repair first does deterministic maintenance/integration repair, then can escalate to whole-repo provider-driven repair if source-level breakage remains
• successful repair only applies a patch back to the root worktree; it does not auto-commit or auto-push

Audit health

npx agent-memory validate

Audits state integrity, history continuity, checkpoint presence, projection alignment, entry wiring, and recall backlog health. It also inspects retention config plus archive batch health under .agent-memory/archive/.

Troubleshooting

Common cases:

• sync reports failed=... The source is still registered, but one or more session files could not be parsed or normalized. Run agent-memory validate to see whether this is now a warning condition.
• recall says Nothing to recall Either there are no unrecalled events for the selected scope, or consolidation produced no durable bundle changes.
• status suggests recall The backlog is above the configured threshold or the checkpoint drift summary suggests stale active memory.
• query says there is not enough evidence The current bundle, history, and checkpoints do not support a confident answer yet. Import more history or run recall before asking again.
• validate warns about recall backlog New history has accumulated and should be consolidated with agent-memory recall.

Breaking Change

This is a destructive model change.

• old state.json formats are not supported
• old projection markers are not supported
• there is no migration path
• old repositories must rerun npx agent-memory init

Learn More

• Overview
• Commands
• File Model
• V1 Boundary And Roadmap
• Adoption Guide
• Contributing

License

MIT