Memory Write Policies

PROJECT CONTEXT:
My agent's memory fills with trivia, duplicates, and things that later turn out false. Memory-system research is blunt: recall cannot fix what careless writing ruined - the write path needs policy, provenance, and validation.

GOAL:
Implement a write policy layer: criteria gates, provenance and confidence metadata, contradiction checks, and write-time validation.

REQUIREMENTS:
- Written criteria for memorable vs not-memorable content
- Every memory entry carries provenance (source, session, date) and confidence
- Contradiction detection: new writes checked against existing memory
- Validation rules per memory type (facts vs preferences vs recipes)
- A quarantine path for low-confidence writes pending confirmation

CONSTRAINTS:
- Treat memory writes like untrusted input - validate, never just append
- When new information contradicts old, resolve explicitly; never store both silently

DEFINITION OF DONE:
- Trivial/ephemeral content is rejected by the gate with a logged reason
- Every stored entry shows provenance and confidence on inspection
- A contradicting write triggers resolution (supersede, reject, or flag), not silent coexistence
- Low-confidence entries live in quarantine until confirmed by a second observation

COMMON FAILURES TO AVOID:
- Remembering everything, so recall drowns in noise
- No provenance, so bad memories cannot be traced or trusted
- Contradictory facts accumulating until the agent's worldview is incoherent
- Single-observation 'facts' treated with the same trust as confirmed ones