ihorponom/agentpack

Repo-native task continuity for AI coding agents.
> Agent memory remembers how you work. Agentpack remembers where the task stands.
When a chat closes or its context is compacted, the next session may need to
reconstruct the task from files and conversation history. Decisions and failed
approaches can be easy to miss.
Agentpack keeps a small, reviewable task ledger in .agentpack/ inside your repo. Connected agents record durable state as they work: the goal, decisions, dead ends, verification evidence, and checkpoints. The next session loads it back and continues. That next session can be the same agent after compaction, a different client, or you returning next week.
Here it is live. In session 1, Claude Code investigates a flaky test and records what it learns through the Agentpack MCP tools. In session 2, the next day with an empty context, the agent loads the task state and picks up where the first session stopped, without re-investigating:
Prefer to try it by hand? The same flow driven from the CLI is in docs/DEMOS.md.
- **Local-first.** Plain files in your repo. No cloud, no telemetry, no network calls.
- **Agent-oriented.** A local MCP server plus generated project instructions (AGENTS.md, CLAUDE.md, Cursor rules) tell agents when to load and record state.
- **Human-friendly.** The same state is available through the CLI for inspection, debugging, and manual handoff.
How is this different from agent memory?
Claude Code memory, Codex memories, and CLAUDE.md or AGENTS.md remember how to work: your preferences, project conventions, and facts that stay true. Agentpack records where the work is: the current task's goal, constraints, decisions, dead ends, next actions, and whether the result was verified.
| | Agent memory | Agentpack |
|---|---|---|
| Holds | Preferences, conventions, long-lived facts | The state of a specific task |
| Lives in | One client's profile or account | .agentpack/ in the repo, local by default, portable with task bundles |
| Read by | That client only | Any MCP client: Claude Code, Codex, Cursor, Claude Desktop |
| Shape | Free-form notes recalled by relevance | A Task Passport with lifecycle, write scope, and verification bound to a commit |
| Staleness | Not tied to file changes | Source conclusions carry file hashes, so changed files are flagged |
Use both: memory for how you like to work, Agentpack for what the task needs next.
Quick start
Requires Node.js >= 20.
In a Git repository, init adds patterns for untracked local files to Git's
info/exclude. It leaves the project's .gitignore untouched. Files already
tracked by Git remain tracked.
Restart or reconnect the coding-agent client. From then on the agent loads Agentpack context at session start, records decisions and evidence while working, and checkpoints meaningful progress.
Run agentpack doctor to verify the setup, and agentpack resume --preset agent --query " " to see the task state yourself.
See docs/INTEGRATIONS.md for client-by-client setup, including what each installer writes and why.
How it works
1. At session start, the agent loads compact Agentpack context: the current Task Passport, recent checkpoints, decisions, and reviewed conclusions about source files.