# Evolution 81 — Memory Evaluation

## Implemented

Encrypted (AES-256-GCM) rows, separate 0600 keyring, individual/specialist/team/mission/org scopes,
provenance, timestamped observations, versioned knowledge, contradiction detection, expiration,
correction/supersession, retrieval confidence, ACL, audit trail, and a promotion guard that forbids
turning an unverified hypothesis into organizational truth without an independent approver and an
evidence reference.

## Constructed suite (16 questions, real model reader)

| arm | correct | n | accuracy | context hits |
|---|---|---|---|---|
| governed_retrieval | 13 | 16 | 81.2% | 22 |
| naive_recent | 1 | 16 | 6.2% | 96 |
| no_memory | 1 | 16 | 6.2% | 0 |
| ungoverned_retrieval | 12 | 16 | 75.0% | 25 |

Deterministic retrieval oracle (no model): {"governed_retrieval": {"accuracy": 0.9375, "correct": 15, "n": 16}, "naive_recent": {"accuracy": 0.125, "correct": 2, "n": 16}, "no_memory": {"accuracy": 0.0625, "correct": 1, "n": 16}, "ungoverned_retrieval": {"accuracy": 0.9375, "correct": 15, "n": 16}}.

## Integrity results

| check | result |
|---|---|
| ciphertext opaque at rest | `True` |
| keyring permissions | `0o600` |
| correction supersedes | `True` |
| self-promotion refused | `True` |
| promotion without evidence refused | `True` |
| expired excluded | `True` |
| contradictions preserved | `0` |

## Public benchmarks

**BLOCKED.** dataset files are not present on this host and downloading them would require an explicit data-acquisition decision
Required contract for a future run: dataset name, SHA-256, license, version, split and command.
No LongMemEval/LoCoMo/MemoryAgentBench score is claimed. G18 is **BLOCKED**.

## Limits

Lexical BM25 retrieval, not embeddings or a vector database. The 16-question constructed suite is
not a public benchmark. The small model's answer accuracy is partly a model limitation.
