---
license: mit
tags:
- ai-memory
- ai-agents
- mcp
- benchmark
- retrieval
- memorygym
---

# Audrey Memory Benchmark Artifacts

This dataset publishes current Audrey benchmark artifacts for inspection and reproducibility.

Project: https://github.com/Evilander/Audrey
Package: https://www.npmjs.com/package/audrey
Live report Space: https://huggingface.co/spaces/Evilander/audrey-memory-benchmark-report

## Current MemoryGym rerun after Audrey tag-isolation fix

Generated: `2026-05-01T04:38:49.849Z`  
Local time: April 30, 2026, 11:38 PM Central  
Suite: `MemoryGym Core`  
Transport: Audrey HTTP sidecar on loopback  
Providers: mock embeddings, mock LLM, isolated data directory

Audrey recall previously treated multiple tag filters as OR. MemoryGym sends `memorygym`, run id, and scenario id together; OR matching let unrelated scenario memories leak through. Audrey now requires all requested tags to match.

| Adapter | Score | Hit rate | P95 recall latency |
|---|---:|---:|---:|
| typed-semantic | 72.7% | 100.0% | 0.4 ms |
| hybrid | 72.7% | 100.0% | 0.3 ms |
| audrey-http | 57.5% | 85.7% | 3.8 ms |

## Audrey before/after

| Run | Score | Hit rate | P95 recall latency |
|---|---:|---:|---:|
| Before tag fix | 34.3% | 42.9% | 3.7 ms |
| After tag fix | 57.5% | 85.7% | 3.8 ms |

## Included files

- `memorygym-before-tag-filter.json`: raw MemoryGym run before the Audrey tag fix
- `memorygym-after-tag-filter.json`: raw MemoryGym run after the Audrey tag fix
- `memorygym-current-run.json`: current raw run, currently same as after-fix run
- `memorygym-current-summary.json`: compact summary and finding notes
- `memorygym-before-after.svg`: before/after graph
- `memorygym-score.svg`: score graph
- `memorygym-hit-rate.svg`: hit-rate graph
- `memorygym-p95-recall-latency.svg`: latency graph
- `audrey-benchmark-report.png`: full-page screenshot of the current report
- `index.html`: static report source

## Claim boundary

This is a current MemoryGym local benchmark artifact. It is not an official AMB,
LoCoMo, LongMemEval, Papers with Code, or CodeSOTA leaderboard score.

The current run is intentionally published with the remaining gaps. The next fix
targets are current-belief precedence and context-aware ranking.