Anamnesis: An Authority-Preserving Runtime Contract for Persistent AI Memory

Persistent AI agents need more than searchable history: transcripts preserve events but not currency, vector indexes find similar passages but not authority, and summaries save tokens while hiding uncertainty and lineage.

Anamnesis is a runtime contract for persistent AI memory. Its central commitment is a five-plane separation: source authority, epistemic status, retrieval utility, visibility authority, and executable authority remain distinct state coordinates through encoding, consolidation, recall, revision, and deletion. Its strongest invariant is authority non-amplification: repetition, summarization, successful retrieval, and favorable outcomes cannot raise authority on their own — a memory transformation may preserve or lower authority, but cannot create it without a separately recorded grant.

The contract further specifies evidence-bearing records, typed navigation associations, bi-temporal validity, bounded associative reconstruction, selective source hydration, lineage-preserving revision, and deletion closure, together with formal rules, proof sketches, an auditable comparison against 22 adjacent memory systems, and a conformance program.

Version 1.0 — conceptual specification. This document makes no generalized performance claim and does not disclose a product implementation. A conformance suite and empirical companion are planned as future versions under this record.