Description

Implement memory retrieval failure tracking as proposed by OmniMem (arXiv:2604.01007, #3566).

Extend skill_outcomes (or add a new memory_retrieval_failures table) to record:

• No-hit turns (query returned 0 results from SYNAPSE spreading activation)
• Low-confidence recalls (score below configurable threshold)
• Query context at failure time: predicate, subject, expected answer, returned answer, latency

This is the minimum viable substrate for future self-improvement experiments. Without a failure dataset, no SYNAPSE parameter tuning experiment can be designed or evaluated. The write path must be async/fire-and-forget to avoid latency impact on the retrieval hot path.

The scheduler-driven SYNAPSE parameter tuning loop described in the OmniMem paper is out of scope for this issue. It will be filed as a separate issue once sufficient failure data has accumulated.

Acceptance Criteria

• SQLite migration adds the required columns or table
• Write path records failures on no-hit and sub-threshold-confidence retrievals
• A sample query in the playbook demonstrates >100 rows after a standard test session
• No measurable latency regression on the retrieval hot path (async write)

References

• OmniMem paper: https://arxiv.org/abs/2604.01007
• Code reference: https://github.com/aiming-lab/OmniMem
• Research synthesis: research(memory): OmniMem autoresearch-guided discovery of lifelong multimodal agent memory — +411% F1 on LoCoMo #3566
• APEX-MEM spec: /specs/004-memory/004-7-memory-apex-magma.md §14.2
• Implementation notes: .local/testing/playbooks/memory-research-2026-04.md