context-management

To prevent context drift, apply a deterministic context-management loop that maintains one active hypothesis and one verification target, ensuring only validated evidence stays in the working set. This stops agents from speculating on stale assumptions and losing their decision trail.

The best way to prepare for compaction or handoff is to generate a five-field checkpoint that captures the active question, verified facts, and a minimal evidence plan. This allows a successor agent to cold start rapidly and rebuild only the necessary context.

Manage context intake by using evidence-first triage to categorize candidate data into must-have, useful soon, durable background, or noise buckets. This prevents large unneeded reads and keeps the working set shaped strictly around the active question.

Selective recovery after context loss works by using the preserved five-field checkpoint to resume the session. It rebuilds only the specific context required by the current active question, rather than reloading the entire previous environment.

You should enforce anti-drift mechanisms whenever your agent session feels noisy, inconsistent, or when the primary verification target changes mid-stream. This ensures unverified assumptions are dropped and the active question doesn't shift unexpectedly.