evanflow-compact

To prevent context drift in long-running AI agent sessions, apply proactive phase-boundary compaction to preserve critical anchors and decisions while discarding irrelevant history, maintaining focus and decision quality across iterations.

The best way to compact context without losing critical anchors is to invoke compaction at clean phase boundaries, which preserves memory anchors like CONTEXT.md and architectural decision records while removing degraded or irrelevant session data.

You can isolate side-quests from your main AI agent context thread by using branch isolation, a technique that moves secondary explorations into separate sessions to prevent pollution of the primary working context.

Context compaction should be applied during major phase boundaries or after lengthy back-and-forth sessions, proactively stabilizing the context to prevent memory degradation before decision quality is impacted.

Yes, session compaction supports preserving architectural decision records across iterations through anchor-based memory preservation, ensuring critical decisions and open questions survive the context compaction process.

A limitation of proactive context compaction is that it requires a clean phase boundary to execute effectively, meaning mid-task invocation may disrupt the active reasoning flow if side-quests are not properly isolated beforehand.