Survive compaction
Keep agent state after context compaction
When a coordinator or long-running agent compacts its context, task status and decisions shouldn't go with it. Keep them in the workspace, not the chat.
The problem
The coordinator knows which lane is done and why, but only in its context window. After compaction it summarises loosely or forgets, and the next turn starts from a guess.
What people try
- Ask the agent to write a status summary before compacting
- You have to remember to ask, and the summary is only as good as what the agent decides to keep.
- A STATUS.md the coordinator keeps updated
- Free-form, easy to skip, and other agents rarely read it before acting.
With Concord
01
Record as you go
Agents write intent, decisions, assumptions, blockers and findings to the task with update_work, while the reasoning is still fresh.
02
Compact freely
That record lives in the repo's local .concord/ store, not in any session, so compaction or a restart loses none of it.
03
Rehydrate
After compaction, inspect_work returns the active claims and the task's recorded memory. The agent picks up from facts, not its own summary.
Result
Compaction stops being a reset. Status and decisions come back from the workspace.
Common questions
- Does this replace /compact or auto-compaction?
- No. Compaction still manages the context window. Concord makes sure the facts that matter to other agents were written somewhere compaction can't touch.
- Where is the state stored?
- In a SQLite database under .concord/ at the repository root. It is local and gitignored by default.
- Will the agent remember to re-read it?
- concord setup writes instructions into CLAUDE.md, AGENTS.md and the other client configs telling agents to call inspect_work when resuming.
Further reading
Give your agents shared work-state.
npm install -g @concord-ai/concord-mcp