Key answer
Continue when goals and valid material remain continuous. Start anew after a handoff when the task changes objects or old assumptions repeatedly interfere. Do not clear by fixed turn count or copy only the final “continue.”
Full explanation
Laban et al. observed dependence on early incomplete answers in simulated multi-turn requirement disclosure. Later experiments compared restatement/summarization and recommended consolidation before retrying. Section 7.4 Anthropic describes compaction and structured notes retaining key decisions, unresolved issues and implementation information. Engineering guide
Preserve what the next step needs, not the entire discussion verbatim:
· Current goal, deliverables, acceptance criteria and changed/withdrawn requirements.
· Valid source locations, versions and evidence for key conclusions rather than only AI summaries.
· Completed artifacts and real verification results, remaining work and failed attempts.
· Decisions, authorized continuation scope and specific unresolved user choices.
For example:
Prepare a handoff for a new conversation using final requirements, current file locations/versions, verified results, remaining work and next action. Remove redundant discussion of rejected approaches but retain key rejection reasons. Separate confirmed facts from unverified assumptions.
Have the new assistant read artifacts and necessary originals, then resume unfinished work. “Tests passed” needs records and “saved” needs an openable file. A new conversation changes the starting point; project resources and cross-chat memory still depend on product/settings.
References
1. Laban et al. (2025), LLMs Get Lost In Multi-Turn Conversation, Section 7.4 and Appendix M
2. Anthropic (2025), Effective context engineering for AI agents, Compaction and Structured note-taking