strategic-compact

Context drift in long AI sessions happens when automated memory compaction discards critical decisions and state, causing the model to lose track of earlier requirements. Strategic compaction prevents this by prompting manual, selective snapshotting at logical phase boundaries to maintain continuity.

You should trigger a context handoff during multi-phase work at major phase transitions, such as moving from research to implementation or before starting a new session, to persist state and guide next steps with sufficient context.

You can configure compaction thresholds using environment variables to tailor the cadence and guardrails of the pre-tool hook script, ensuring large-output tasks are spawned with sufficient context before major phase transitions.

Yes, you can automate context snapshots using a lightweight pre-agent compaction hook script that checks environment-configurable thresholds and ensures large-output tasks are spawned with sufficient context before execution.

The best way to manage memory drift across multiple phases is enforcing a defined handoff protocol with manual, boundary-aligned compaction prompts, which persists state and guides next steps to maintain continuity across transitions.

Automated context compaction erodes critical session state because it is a lossy process that indiscriminately removes information, whereas strategic compaction advocates manual, boundary-aligned snapshotting to preserve fidelity.