# Evolution 80 — Memory Evaluation

> Generated by `scripts/evolution80/docs_e80.py` from `artifacts/e80/` (run 3). FIRME AI CORPORATION — a three-person team (founder: Vincent Noriega).
> Development branch `evolution-80`. Not deployed. No cloud model was called. No third-party audit.

Answer accuracy with a real local model (qwen2.5-0.5b-instruct), four arms, constructed question set. Retrieval hit rate is
reported separately from answer accuracy: storing or retrieving a record is not proof of a correct answer.

| arm | correct | answer accuracy | Wilson 95% | retrieval hit rate (answerable) | poison in context | answered with poison | tokens |
|---|---|---|---|---|---|---|---|
| governed_retrieval | 28/28 | 100.0% | [87.9%, 100.0%] | 100.0% | 0 | 0 | 3133 |
| naive_recent | 6/28 | 21.4% | [10.2%, 39.5%] | 0.0% | 0 | 0 | 5197 |
| no_memory | 6/28 | 21.4% | [10.2%, 39.5%] | 0.0% | 0 | 0 | 2285 |
| ungoverned_retrieval | 20/28 | 71.4% | [52.9%, 84.8%] | 72.7% | 13 | 2 | 3767 |

| comparison | discordant pairs | p (two-sided) | significant at 0.05 |
|---|---|---|---|
| governed_vs_no_memory | 22 | 0.0 | True |
| governed_vs_naive | 22 | 0.0 | True |
| governed_vs_ungoverned | 8 | 0.007812 | True |

## Store

71 records across 10 memory types; 3 quarantined; 2 superseded;
hash chain intact: True; plaintext absent at rest: True;
every hit carries `grants_authority: false`: True.

## Public benchmarks

| benchmark | status |
|---|---|
| AMA-Bench | NOT_EVALUATED (dataset not run; context lengths exceed what a 0.5B CPU model can use within budget) |
| LoCoMo | NOT_EVALUATED (dataset not run; context lengths exceed what a 0.5B CPU model can use within budget) |
| LongMemEval | NOT_EVALUATED (dataset not run; context lengths exceed what a 0.5B CPU model can use within budget) |
| MemoryAgentBench | NOT_EVALUATED (dataset not run; context lengths exceed what a 0.5B CPU model can use within budget) |

## Limits

Constructed set, 30 questions, lexical BM25 retrieval, one small model. Retrieval hit-rate is not reasoning accuracy; both are shown. Injection screening is a regex and misses paraphrases. Test-time learning was not measured.
