Memory
Operator insight (2026-10-08): the AI memory product is part of the surface. Two roles in one surface: (1) maintained-by-us structured knowledge — obligations, rules, classifications (canon-derived, versioned) served for models to use; (2) day-to-day user memories. KEY separation: memory does NOT decide — it provides structured information; decisions live in engine/gates. Consequences: two trust tiers (vendor-maintained signed content vs user memories, different write paths/provenance, one retrieval surface); poisoning memory yields at worst bad advice, never unauthorized action (gates enforce independently); maintained-memory currency IS the subscription shape (we keep knowledge current); provenance pipeline feeds maintained memory directly; R0 scopes which maintained content applies per deployment. [Digest: 53df278b-50a4-44ed-a3d2-156dfa06cc37]
Tags: memory-product, architecture, separation-of-powers, project:redaction-kernel, kind:semantic, rolled-up — Source: mcp — 2026-10-08 14:12:39 UTC