PROJECT: memory-compactor

Aim

Compact noisy conversational memory stretches (especially debug sessions) into durable resolved summaries while preserving recall quality.

Current implementation

HTTP client for custom mem0 server endpoints:
- POST /memories/all
- POST /memories/raw
- DELETE /memory/{id}
temporal clustering (gap-minutes)
LLM segment extraction + summary writer
dry-run reporting

Operating mode

default dry-run
--apply to mutate memory state
bounded deletions with --max-deletes

Design notes

write compacted summary before deleting old entries
append extracted facts into the summary text (non-empty only)
annotate summaries with provenance metadata (segment_source_ids, extraction)
set metadata.created_at to the segment end timestamp so summary ordering stays intact

Next steps

Add rollback snapshot file before apply
Add explicit per-user allowlist
Add cron integration helper (OpenClaw cron payload template)
Add quality checks pre/post compaction