roadmap: make BrainLayer self-maintaining

[EtanHey]

Ruling

BrainLayer must maintain its own memory quality. Deduplication, supersession, correction, enrichment, and hygiene cannot depend on agents remembering to invoke maintenance tools.

The July 2026 nine-tool profile incident demonstrated the failure mode: maintenance calls were unavailable for days while stale memories continued to win retrieval, with no autonomous repair loop.

Scope

Design a background maintenance subsystem that:

• detects duplicate, stale, contradicted, and under-enriched memories;
• schedules enrichment, deduplication, correction, supersession, and archival work without agent prompting;
• preserves the one-writer contract and explicitly handles SQLite locking, WAL growth, retries, and backpressure;
• records auditable decisions and exposes health/progress metrics;
• applies conservative safety gates for personal data and user-intent memories;
• fails visibly rather than silently degrading memory hygiene.

Non-goal for the urgent nine-tool unblock

Do not build this system in the gated-tool error/runtime-unblock patch. This issue is the durable roadmap record for a separately designed and reviewed project.

Definition of ready

• architecture and failure-mode review completed;
• maintenance policy and safety boundaries documented;
• measurable quality/health criteria defined;
• rollout plan includes shadow mode, observability, recovery, and WAL/lock verification.