Retrieving evidence again does not guarantee recovering an agent’s decision history after compaction.

Anthropic recommends lightweight references for dynamic retrieval and structured notes for long tasks, warning that aggressive compaction can discard important details. The uncertainty is whether the agent resumes with the reasoning behind prior decisions, not merely access to their sources.

For a startup pilot, I recommend a context map separating evidence to fetch from task state to preserve. Keep decisions, their rationale, source references and unresolved work in that record. This gives the agent a basis for continuing or reconsidering, rather than silently replacing earlier reasoning.

My proposed test would trigger compaction after a decision is recorded but before related work is resolved. Compare the resumed work against a saved pre-compaction record. Treat forgotten constraints, dropped work and unexplained reversals as failures; accept changed decisions when the agent preserves their history and explains the change with evidence.

Runtime exploration may be slower than precomputed retrieval. I would compare elapsed time and token use against an upfront-context baseline on identical tasks, separately from continuity checks. Choose retrieval timing based on that cost, without accepting lost task state as an efficiency gain.

#ContextEngineering

#AIAgents

Which decision history must your agent preserve so a justified revision remains distinguishable from forgetting?