| <!doctype html> |
| <html lang="en"> |
| <head> |
| <meta charset="utf-8" /> |
| <meta name="viewport" content="width=device-width, initial-scale=1" /> |
| <title>Audrey MemoryGym Current Benchmark Evidence</title> |
| <style> |
| :root { color-scheme: light; --ink:#111827; --muted:#4b5563; --line:#d7dee8; --soft:#f5f7fb; --blue:#1d4ed8; --teal:#0f766e; --amber:#b45309; --red:#b91c1c; } |
| * { box-sizing: border-box; } |
| body { margin:0; font-family: Inter, ui-sans-serif, system-ui, -apple-system, Segoe UI, Arial, sans-serif; color:var(--ink); background:#fff; } |
| header { padding:52px 28px 28px; border-bottom:1px solid var(--line); background:linear-gradient(180deg,#f8fbff 0%,#fff 88%); } |
| main { max-width:1120px; margin:0 auto; padding:28px; } |
| .wrap { max-width:1120px; margin:0 auto; } |
| h1 { margin:0 0 10px; font-size:44px; line-height:1.05; letter-spacing:0; } |
| h2 { margin:34px 0 14px; font-size:24px; } |
| h3 { margin:0 0 10px; font-size:18px; } |
| p { line-height:1.55; color:var(--muted); max-width:880px; } |
| .lede { font-size:18px; color:#263241; } |
| .stamp { display:inline-flex; gap:8px; align-items:center; padding:7px 10px; border:1px solid var(--line); border-radius:8px; background:#fff; font-size:14px; color:#374151; margin-bottom:18px; } |
| .grid { display:grid; grid-template-columns:repeat(3,minmax(0,1fr)); gap:16px; margin:24px 0; } |
| .card { border:1px solid var(--line); border-radius:8px; padding:18px; background:#fff; } |
| .metric { font-size:34px; font-weight:800; margin:2px 0 4px; } |
| .bad { color:var(--red); } |
| .ok { color:var(--teal); } |
| .charts { display:grid; grid-template-columns:1fr; gap:18px; } |
| .charts img { width:100%; border:1px solid var(--line); border-radius:8px; background:white; } |
| table { width:100%; border-collapse:collapse; margin:12px 0 22px; border:1px solid var(--line); border-radius:8px; overflow:hidden; font-size:14px; } |
| th, td { padding:12px 13px; border-bottom:1px solid var(--line); text-align:left; vertical-align:top; } |
| th { background:var(--soft); font-size:13px; color:#374151; } |
| tr:last-child td { border-bottom:0; } |
| code { background:#eef2f7; border:1px solid #d8e0eb; border-radius:5px; padding:2px 5px; font-size:0.92em; } |
| .callout { border-left:4px solid var(--amber); background:#fff8e7; padding:14px 16px; border-radius:6px; margin:18px 0; } |
| .done { border-left-color:var(--teal); background:#eefcf7; } |
| .files { columns:2; padding-left:18px; color:#374151; } |
| a { color:#174ea6; } |
| @media (max-width:760px) { h1{font-size:34px;} .grid{grid-template-columns:1fr;} .files{columns:1;} main, header{padding-left:18px;padding-right:18px;} } |
| </style> |
| </head> |
| <body> |
| <header> |
| <div class="wrap"> |
| <div class="stamp">Current rerun generated 2026-05-01T04:38:49.849Z (Apr 30, 2026, 11:38 PM Central)</div> |
| <h1>Audrey MemoryGym Benchmark Evidence</h1> |
| <p class="lede">This page shows the current MemoryGym Core rerun after an Audrey isolation fix. It replaces the stale synthetic Audrey chart and includes the before/after numbers.</p> |
| <div class="callout"><strong>Claim boundary:</strong> this is a current MemoryGym local benchmark artifact. It is not an official AMB, LoCoMo, LongMemEval, Papers with Code, or CodeSOTA leaderboard score.</div> |
| </div> |
| </header> |
| <main> |
| <section class="grid" aria-label="Audrey run summary"> |
| <div class="card"><h3>Audrey HTTP Score</h3><div class="metric bad">57.5%</div><p>Up from 34.3% before the tag isolation fix (+23.2 pts).</p></div> |
| <div class="card"><h3>Audrey Hit Rate</h3><div class="metric ok">85.7%</div><p>Expected memory appeared in top-k on 6 of 7 probes, up from 3 of 7.</p></div> |
| <div class="card"><h3>Audrey P95 Recall</h3><div class="metric ok">3.8 ms</div><p>Latency changed +0.1 ms after the fix.</p></div> |
| </section> |
|
|
| <h2>Before And After</h2> |
| <div class="charts"><img src="memorygym-before-after.svg" alt="Audrey MemoryGym before and after chart" /></div> |
|
|
| <h2>Current Graphs</h2> |
| <div class="charts"> |
| <img src="memorygym-score.svg" alt="MemoryGym score chart" /> |
| <img src="memorygym-hit-rate.svg" alt="MemoryGym hit rate chart" /> |
| <img src="memorygym-p95-recall-latency.svg" alt="MemoryGym p95 recall latency chart" /> |
| </div> |
|
|
| <h2>Adapter Results After Fix</h2> |
| <table><thead><tr><th>Adapter</th><th>Score</th><th>Hit Rate</th><th>P95 Recall</th><th>Precision</th><th>Contamination Penalty</th></tr></thead><tbody><tr><td>typed-semantic</td><td>72.7%</td><td>100.0%</td><td>0.4 ms</td><td>34.5%</td><td>45.2%</td></tr> |
| <tr><td>hybrid</td><td>72.7%</td><td>100.0%</td><td>0.3 ms</td><td>34.5%</td><td>45.2%</td></tr> |
| <tr><td>audrey-http</td><td>57.5%</td><td>85.7%</td><td>3.8 ms</td><td>29.8%</td><td>50.0%</td></tr></tbody></table> |
|
|
| <h2>What Changed</h2> |
| <p>Audrey recall previously treated multiple tags as an OR filter. MemoryGym passes <code>memorygym</code>, the run id, and the scenario id on recall. OR matching let unrelated scenario memories through because every row shared the <code>memorygym</code> tag. Audrey now requires all requested tags to match, which raised Audrey HTTP from 34.3% to 57.5%.</p> |
|
|
| <h2>Remaining Failures</h2> |
| <p>The next benchmark work is current-belief precedence and context-aware ranking. Audrey now usually finds the expected memory, but still ranks stale or wrong-project memories too high.</p> |
| <table><thead><tr><th>Probe</th><th>Query</th><th>Result</th><th>Returned IDs</th><th>Top Recall</th></tr></thead><tbody><tr><td>typed-profile-updates / current-workspace</td><td>Which collaboration environment should Maya use right now?</td><td>0.0% score<br>0.0% hit</td><td>maya-workspace-old, jonas-distractor-workspace, maya-pref-morning</td><td>Maya's active workspace used to be Atlas for the onboarding sprint.</td></tr> |
| <tr><td>typed-profile-updates / meeting-preference</td><td>How should Maya receive meeting preparation?</td><td>70.7% score<br>100.0% hit</td><td>maya-pref-morning, maya-workspace-old, maya-workspace-current</td><td>Maya prefers morning async notes before any meeting-heavy work.</td></tr> |
| <tr><td>retrieval-context-routing / audrey-release-gate</td><td>Which project requires a benchmark gate and doctor check for release validation?</td><td>59.0% score<br>100.0% hit</td><td>termivibe-first-contact, audrey-benchmark-gate, repopulse-trace-contract</td><td>TermiVibe first contact validation should cover config precedence, typed mode, no wake word mode, and telemetry persistence.</td></tr> |
| <tr><td>retrieval-context-routing / repopulse-trace</td><td>Which memory mentions pipeline version and dense rerank trace boundaries?</td><td>57.0% score<br>100.0% hit</td><td>audrey-benchmark-gate, repopulse-trace-contract, termivibe-first-contact</td><td>Audrey release validation includes build, typecheck, benchmark gate, doctor, demo, pack dry-run, and host smoke coverage.</td></tr></tbody></table> |
|
|
| <h2>Submission Status</h2> |
| <div class="callout done"><strong>Already submitted/listed for review:</strong> MCP.Directory, MCP.so, Glama, CodeSOTA coverage request, Hugging Face dataset/Space, and AMB maintainer issue for the official provider path.</div> |
| <table> |
| <thead><tr><th>Target</th><th>Status</th><th>URL / next action</th></tr></thead> |
| <tbody> |
| <tr><td>MCP.Directory</td><td>Submitted for review</td><td>Review promised within 24 hours.</td></tr> |
| <tr><td>MCP.so</td><td>Signed-in server record created and GitHub issue opened</td><td><a href="https://github.com/chatmcp/mcpso/issues/2198">mcpso issue 2198</a></td></tr> |
| <tr><td>Glama</td><td>Submitted for review</td><td>Signed-in form completed.</td></tr> |
| <tr><td>CodeSOTA</td><td>Coverage request submitted, not a fake score</td><td>Waiting for editorial reply.</td></tr> |
| <tr><td>AMB</td><td>Provider/leaderboard issue opened and updated with this run</td><td><a href="https://github.com/vectorize-io/agent-memory-benchmark/issues/11">AMB issue 11</a></td></tr> |
| <tr><td>Hugging Face</td><td>Dataset and Space updated with current evidence</td><td><a href="https://huggingface.co/spaces/Evilander/audrey-memory-benchmark-report">Space</a> and <a href="https://huggingface.co/datasets/Evilander/audrey-memory-benchmark-artifacts">Dataset</a></td></tr> |
| </tbody> |
| </table> |
|
|
| <h2>Files In This Evidence Bundle</h2> |
| <ul class="files"> |
| <li><code>memorygym-before-tag-filter.json</code></li> |
| <li><code>memorygym-after-tag-filter.json</code></li> |
| <li><code>memorygym-current-run.json</code></li> |
| <li><code>memorygym-current-summary.json</code></li> |
| <li><code>memorygym-before-after.svg</code></li> |
| <li><code>memorygym-score.svg</code></li> |
| <li><code>memorygym-hit-rate.svg</code></li> |
| <li><code>memorygym-p95-recall-latency.svg</code></li> |
| <li><code>audrey-benchmark-report.png</code></li> |
| </ul> |
| </main> |
| </body> |
| </html> |
