Give different kinds of memory different homes
Long conversations mix policy, temporary observations, source material, and unfinished work. Separate those concerns. Durable instructions describe how the job works; task state records what is happening now; evidence supports the output.
A memory file is still input to the agent. It can become stale or contain an incorrect conclusion. Keep an owner, scope, and review process rather than assuming that everything saved should influence every future run.
- Instructions: versioned workflow contract and approved conventions.
- State: current run ID, completed stages, blockers, and next action.
- Evidence: sources, outputs, and validation results.
- Decisions: why a choice was made and when it should be revisited.
Write a handoff another operator can use
A handoff should let a person or another agent continue without reconstructing the full transcript. Name the goal, last verified checkpoint, exact files, outstanding questions, and permission boundary.
Distinguish verified facts from working assumptions. Include commands or actions already attempted so the next worker does not repeat an external action just because the conversation history is short.
Goal: prepare the weekly briefing draft
Verified: sources collected; links checked
Artifacts: sources.json, drafts/briefing.md
Blocked: two sources disagree on launch date
Next: reviewer resolves conflict before approval
Do not: publish, email, or modify the source list
External actions already completed: noneKeep sensitive material out of reusable memory
Use references to approved secret storage instead of pasting credentials into notes. Avoid carrying customer records or private messages into a general-purpose skill. Keep only the information needed for the job and follow your organization’s retention rules.
Review memory when a workflow or source changes. Archive superseded decisions and remove misleading instructions. More memory is not automatically better context.
Test continuation as a real workflow
Pause after a checkpoint and ask a new run to continue from the handoff alone. Check whether it finds the right artifacts, respects the boundary, and identifies unresolved questions. Record missing context as a test failure.
Use the same exercise after a harness update or a change to storage paths. A handoff is useful only if it works outside the session that created it.
Sources and further reading
These guides combine linked documentation with our suggested operating practices. Examples are illustrative; they are not customer results or harness benchmarks.