A serious agent should not treat chat transcript as its source of truth. It needs a portable memory passport: preferences, project decisions, unresolved tasks, research, policies, errors, handoffs, and proof anchors.
What goes into the passport
- Stable user and project preferences.
- Accepted decisions with provenance and timestamps.
- Open tasks and handoff state.
- Research crystals separated from conjecture.
- Ledger proofs and exportable evidence.
Reference agent is the flagship pattern: a bot that remembers, but can also prove why that memory should be trusted.
Why this is commercial
Every AI assistant pitch says it remembers. Very few can answer: who admitted this memory, under what rule, and can I export proof?
The operating model
The passport is not a giant prompt. It is a typed, scoped, aging set of facts that can be retrieved, exported, and challenged. That is the difference between personalization and operational continuity.