Auditable Memory Records

Memory records you can check, not just trust — provenance and epistemic status for AI agents.

Most agent memory systems store what was remembered. Almost none store how strongly it is believed, what evidence supports it, or where exactly it came from — so a stored memory can only be trusted, never checked.

Auditable Memory Records (AMR) specifies a small set of fields that make a memory record checkable:

• auditable_memory — the conformance declaration. Required, and the reason adoption is findable.
• epistemic — is this observed, derived, unresolved, or merely asserted?
• backed_by / contradicts — declared, typed links to supporting and conflicting records
• sources / claims — quote-level citations binding an individual claim to the exact source words that justify it, with a verifiable hash

auditable_memory: "0.1"
epistemic: inference
claims:
  - text: "insulin list prices outpaced general medical inflation"
    source_id: sources/kff-report-2026.md
    span:
      quote: "list prices for insulin products rose 11% between 2024 and 2026"
      quote_hash: "sha256:029a3c30f5e0939075dea1e1e64c18a3b108f2e724c2f0a4b85d1bd7d267e48a"
contradicts: [research/pharma-pricing-brief-2025]

→ Read the specification · version 0.1.0 (draft) · CC BY 4.0

These are not new ideas. Nanopublications, micropublications, PROV-O, TEI and W3C Web Annotation covered most of this ground years ago, some of it far more expressively. AMR's contribution is packaging and audience — a lightweight, non-RDF format aimed at LLM agent-memory implementers. SPEC §10 says exactly what is borrowed.

---

Conformance in one table

Level	Name	What it requires
1	Marked	Emit auditable_memory, preserve epistemic from a closed four-value vocabulary, and never treat absence as fact.
2	Linked	Declared backed_by / contradicts, queryable as relations, never auto-resolved.
3	Cited	Claims bind to source spans with verifiable hashes, and a third party can check them without trusting you.

Level 1 should take an afternoon — one declaration, one preserved field, one rule about absence, and no storage-layer changes. It is deliberately cheap, but not satisfiable by doing nothing: an implementation that emits no auditable_memory field conforms at no level. That declaration is also what makes adoption countable; without it, a record implementing this perfectly is indistinguishable from one that merely looks similar.

Running the conformance suite

conformance/ holds the test vectors, as data rather than code, so any implementation in any language can run them:

File	Covers
normalize.yaml	Quote normalization (SPEC §4) — punctuation folding, whitespace collapse, idempotence
level1-marked.yaml	The closed vocabulary and the absent-is-not-fact rule
level2-linked.yaml	Ref hygiene, traversal rejection, non-resolving contradicts
level3-cited.yaml	The four verification outcomes, hash prefixing, claim_id derivation

Each case carries input, expect, and why. Cases marked THE LOAD-BEARING CASE in their why are the ones that distinguish real conformance from field-copying — if you implement nothing else from this suite, implement those.

Every hash in level3-cited.yaml is a real sha256 of the normalized quote and can be recomputed independently:

hashlib.sha256(normalize(quote).encode()).hexdigest()

Status

This is a 0.1 draft published to find implementers, not a finished standard. The fields are extracted from a working implementation rather than designed on paper — which usefully constrains the design, but also means it currently reflects one system's experience.

The most useful feedback, in order:

1. Attempt Level 1 in your system and report what was awkward.
2. Argue with the closed vocabulary. Four values is a bet. If a fifth is load-bearing for you, that matters more before the vocabulary hardens than after.
3. Break the conformance suite. Cases it misses are the most valuable contribution available.

Relationship to Palinode

Palinode is the reference implementation. It is not a dependency: adopting this specification does not require adopting Palinode, and adoption by other systems is the explicit goal rather than a side effect.

Writing this specification found a real bug in the reference implementation: quote hashes were emitted as bare MD5 digests, which record no algorithm and therefore cannot be migrated. That is now fixed (SPEC §4), and it is the clearest argument available for extracting a specification from working code — the field had been chosen incidentally, and nothing had ever forced it to be justified.

Licence

Specification text: CC BY 4.0. Conformance fixtures: MIT. Both chosen so you can copy them into your own docs and test suites without asking.