# E114 memory architecture (Evidence-Native Memory 3.0)

Layers: **A source** (immutable, content-addressed, UPDATE/DELETE blocked by SQLite triggers) · **B semantic** · **C episodic** · **D causal**
(cause, effect, intervention, confidence required) · **E procedural** (requires a verification source) · **F collective** · **G strategic**.

Write path: `propose` (author) → `validate` (a different identity) → `commit` (new immutable version; the head moves; journal entry) → `revert`/`retire` (head moves; history kept).

Validation refuses: self-validation, unknown layer, records claiming authority, missing evidence, broken references, source-hash mismatch,
injection patterns, **numbers in a claim that appear in no cited source**, semantic/causal/procedural facts supported only by untrusted sources,
incomplete causal edges, and procedures without verification. A changed value on the same key without `supersedes` becomes **CONTESTED**:
contradictions are recorded, never silently resolved. `as_of` queries give temporal reasoning over valid-from times.

Integrity audit: source hashes, record digests, journal chain, and "every record row appears in the journal" (catches inserts that bypass the triggers).

## Measured in the run
- Stats: 93 sources, 8 record versions, rejections {'FABRICATED_NUMBER:24': 1, 'UNTRUSTED_ONLY_EVIDENCE': 1}, integrity True.
- Poisoning: rejected with `UNTRUSTED_ONLY_EVIDENCE`. Behavioural effect of the poison on an agent without validation: identification
  0.692 vs 0.967.
- Behaviour change from memory: sealed REJECTED (identification loss), near transfer CONFIRMED (see gate G5).
- Self-organisation: 0 topic-index proposals staged, none applied.
- Unit tests cover immutability, self-validation, fabricated numbers, broken references, injection, untrusted-only, versioning, contested values, revert, tamper and cross-session persistence.

Not run: AMA-Bench, LongMemEval, LoCoMo (official datasets and LLM-judge protocols were not available offline). No equivalence is claimed.
