Roadmap

Current Capability: L3 → L2 Re-Distillation

Manual re-distillation (POST /v1/mems/redistill or python -m mems.redistill) is a current capability, not a future item: archived JSONL is re-distilled with the current OPENAI_MODEL, deduplicated by content fingerprint + model, and upgraded through the L2 version chain.

Planned: L1 Deletion Strategy

Today Mems never deletes L1 rows — the lifecycle is archive (mark is_archived). If online L1 deletion is added later, it must stay compatible with the re-distillation loop:

• Only delete archived L1: allow deletion only for records already archived (is_archived=True), or archive first, so L3 always keeps a re-distillable copy. Deleting before archiving loses the memory irreversibly and breaks the loop.
• Cascade Qdrant cleanup: deleting an L1 row must also delete its vector replica (delete_points) to avoid orphaned points that waste storage and produce dead candidates.
• Redirect L2 evidence: source_l1_ids pointing to a deleted L1 row become dangling references; the re-distill progress table already records l3_path/l3_line as a fallback evidence pointer, and new L2 lineage should prefer the L3 pointer when the L1 row is gone.

Other Ideas

• Distributed distillation lock (Redis SETNX) so multi-replica deployments do not re-distill the same batch concurrently.
• L0 output pipeline consumption metrics (XLEN / consumer lag) in /status.
• Recall score decomposition (explain) for tuning the hybrid ranker.