Memory model

UnifiedMemory treats memory as governed layered state, not a single vector collection.

Layers

• Core memory: stable canonical facts safe for assistant context.
• Working memory: short-lived task/session context scoped to an agent or run.
• Observations: synthesized patterns that require support before promotion.
• Mental models: narrow operating models activated only after clean evidence gates.
• Archival memory: historical records and raw evidence.
• Forensic residue: test, smoke, benchmark, diag, eval, junk, secret-like, and troubleshooting rows.

Only clean assistant-visible facts, approved observations, and active mental models can enter normal prompts.