Key answer
Inspect how current artifacts use early agreements and compare with original records. “Do you remember?” produces a self-report only.
Full explanation
*Lost in the Middle* found position effects in long-context retrieval. *LLMs Get Lost In Multi-Turn Conversation* split complete requirements over simulated conversations and observed degradation. In the latter's four task types with GPT-4o and GPT-4o-mini, consolidated restatement helped without restoring single-turn performance. These concern long-material use and dispersed requirements, not identical failure in every long conversation. Long-context study, multi-turn study, Section 7.1
At chapter, batch or feature completion, select consequential early agreements: term meanings, source scope, content not to delete and completion criteria. Ask where new artifacts satisfy them, then inspect those locations. Checking application is closer to the goal than reciting every rule.
Check this batch against our first three agreements: retain question IDs, support factual judgments, and do not count unverified work as complete. Identify artifact locations and discrepancies; correct them and continue. If I explicitly changed a condition, list its current replacement too.
For omissions, supply the relevant original or current rule and identify obsolete agreements. After multiple revisions, maintain a versioned current list rather than repeatedly pasting contradictory old and new requirements.
For checkable completion, sources and numbers, inspect actual text, data or tests. Restating rules can reveal misunderstanding but cannot replace output inspection.
References
1. Liu et al. (2024), Lost in the Middle: How Language Models Use Long Contexts, TACL
2. Laban et al. (2025), LLMs Get Lost In Multi-Turn Conversation, study v1, Section 7.1 and Table 2