MS4PI recall and checkpointing

Recall and checkpointing in MS4PI

MS4PI uses layered continuity rather than treating one feature as “memory.”

The main layers are:

• first-person identity and user context
• append-only LOG checkpoints
• Markdown memory files with MEMORY.md index pointers
• archived Pi session JSONL transcripts
• vector-backed recall when embeddings are available
• rich .handoff.md compaction handoff
• checkpoint-bundle approval discipline

Recall

MS4PI indexes memory files and archived Pi transcripts into a local vector store when embeddings are configured. It can fall back to degraded/text-search behavior where semantic embedding support is unavailable.

Useful commands:

/ms-recall-status
/ms-recall-backfill
/ms-recall-search <query>

Checkpoint bundle discipline

A checkpoint is not complete just because a LOG entry exists.

When new or updated memory is warranted, checkpoint completion requires:

1. approved memory file writes/updates
2. approved MEMORY.md index updates
3. approved LOG.md append
4. transcript archive/backfill
5. recall/embed verification

For /ms-checkpoint, one checkpoint-bundle approval covers the approved LOG entry plus warranted memory/index writes unless the human explicitly narrows approval.

Handoff and compaction

.handoff.md carries compacted session continuity across context reduction. The handoff should preserve:

• current state
• decisions
• open work
• verified commands
• non-claims
• drift or risk flags
• next-step guidance

MS4PI should not claim compaction parity beyond what has been validated on the target Pi runtime.