Key answer
A handoff should start the next phase from the right state, preserving decisions, evidence, conditions and unfinished work rather than merely summarizing discussion. Retain source locations for disputed or verbatim requirements.
Full explanation
Anthropic's compaction practice emphasizes key decisions, unresolved issues and implementation details, warning that excessive compression can lose information whose importance appears later. LongMemEval likewise found that replacing original sessions with extracted summaries/facts reduced overall QA performance through information loss in its retrieval-memory setting. A summary is an entry point; key evidence should remain traceable. Engineering explanation, Section 5.2
A practical handoff can use this order, combining items when simple:
1. Goal and acceptance: deliverable, audience and mandatory conditions.
2. Latest artifact: version, location, completion stage and checks performed.
3. Confirmed decisions and reasons: who decided, why, evidence locations and tradeoffs affecting later choices.
4. Exceptions and obsolete items: temporary requirements, superseded rules, rejected approaches and key reasons.
5. Unresolved issues: distinguish missing information, unverified claims and pending user decisions rather than calling all “future optimization.”
6. Next action: resumption point, dependencies and still-valid permissions/limits.
“Refund rules confirmed” is vague. Better: “Eligibility to apply confirmed, application not yet accepted; based on rule 3. No arrival-time promise; review outcome still missing.”
Prepare a new-conversation handoff retaining confirmed decisions and evidence, temporary exceptions, obsolete rules and unfinished work. Separate user decisions from AI suggestions. Preserve key numbers, negations and conditions; link the latest files and original locations.
Before saving, check numbers, scope, state and evidence links. Items labeled verified need records; without evidence, mark them for checking. Date or version updates so a new conversation does not use stale state.
References
1. Anthropic (2025), Effective context engineering for AI agents, Compaction