Memory between runs
Persist agent decisions between sessions
Where does an agent's memory live between runs? Keep decisions, assumptions and "done" status attached to the task, where the next session can read them.
The problem
A session ends and its reasoning ends with it. The next one re-asks settled questions, or trusts a "done" that only ever existed in chat.
What people try
- AGENTS.md or CLAUDE.md notes
- Good for standing rules. Poor for per-task decisions, which pile up and go stale with nothing to say when.
- Re-reading the old chat transcript
- Slow, and other harnesses can't read it at all.
With Concord
01
Type the memory
update_work records decisions, assumptions, questions and blockers as typed entries on the task, not loose prose.
02
Finish with evidence
finish_work records what changed, which tests ran and what's risky, so "done" means something another agent can check.
03
Start from the record
The next session, in any harness, reads the task with inspect_work. Concord also writes HANDOFF.md if a human wants to read it.
Result
Sessions end. The task's memory doesn't, and no harness owns it.
Common questions
- Is this a vector memory store?
- No. It is structured task state: who owns what, and the typed decisions and evidence attached to each task. Nothing is embedded or retrieved by similarity.
- Can I commit the handoff to the PR?
- Yes. .concord/ ignores itself by default, but you can add negation rules to .concord/.gitignore or force-add HANDOFF.md and REVIEW_PACKET.md.
Further reading
Give your agents shared work-state.
npm install -g @concord-ai/concord-mcp