mnemos

To prevent context compaction from losing task memory, you can use a structured memory graph that stores typed nodes for goals, constraints, and results. This ensures critical working knowledge survives compaction and can be re-injected when the session resumes.

Context compaction is a process that compresses active session context to manage token limits. It can destroy structured task knowledge by evicting critical details like goals and constraints unless a typed memory store actively preserves and restores them post-compaction.

You can restore agent hooks and task checkpoints after a session transition by loading persisted checkpoint files from a designated memory directory. A three-layer recovery system re-injects this context at pre-compaction, session-start, and pre-tool fallback stages.

Yes, you can maintain task goals and constraints across multiple Claude Code sessions by managing a typed MnemoGraph memory store. Goals and constraints are never evicted during compaction, ensuring intent and invariants remain consistent across transitions.

Yes, this approach supports iCPG integration by bridging iCPG ReasonNodes directly into the memory graph as Goal and Constraint nodes. This keeps agent reasoning and task invariants aligned and persistent across different working sessions.

Fatigue-scored consolidation for memory checkpointing relies on monitoring token utilization, scope scatter, re-read ratio, and error density. It is limited by the accuracy of these passive metrics in deciding exactly when to compress evictable context nodes or persist checkpoints.