A React website for the paper User as Code: Executable Memory for Personalized Agents. It explains the mechanism and lets you interactively explore every graded test case across all four benchmarks (LOCOMO, LongMemEval, Analytical Inference, Active Service).
- Vite + React 18
- Tailwind CSS
- Recharts (interactive charts)
- Framer Motion (animation)
npm install
npm run dev # http://localhost:5173npm run build # → dist/ (fully static, deployable anywhere)
npm run previewThe site is driven by compact JSON bundles in public/data/, generated from the raw
experiment results by:
python3 ../scripts/build_site_data.pyThis reads experiments/results/*.json, the source benchmarks, and the active-service
scenario definitions, and writes:
| file | contents |
|---|---|
locomo.json |
600 LOCOMO QAs × 7 systems, with conversation-evidence context, dual-judge grades |
longmemeval.json |
500 LongMemEval QAs × 7 systems, dual-judge grades |
analytical.json |
100 analytical cases × 5 systems, with executed Python traces |
active.json |
60 Active-Service scenarios with multi-session context + per-system runs |
summary.json |
aggregate tables / chart data (from the paper) |
Re-run the script if the underlying results change, then rebuild.
- (a) the benchmark result / gold answer
- (b) the context (conversation evidence, seeded sessions, or record set)
- (c) each system's response (including UaC's executed code trace)
- (d) the grading — dual-judged under Gemini 3 Flash and Claude Opus 4.7 where applicable