[Program] Adaptive memory and learning

[francip]

Primary source

Program outcome

Connect durable memory, live adaptation, corrected-data training, evaluation, and bounded serving across memory and agent repositories.

Operating rules

• Keep data preparation, adapter generation, training, evaluation, promotion, and rollback as separate verifiable packets.
• Require privacy review, deterministic splits, artifact provenance, regression gates, and explicit human promotion.
• Do not place private source rows or model artifacts in issue bodies or repositories.