MEMOIR: open-source benchmark tests AI memory on 117 synthetic oncology patients
jefrankle · x · 2026-09-24
MEMOIR (Memory Evaluation over Multi-year Oncology Inference & Retrieval) is a new open-source benchmark for longitudinal clinical memory: 117 synthetic oncology patients, 65,377 clinical events, and 3,617 questions covering factual retrieval, rare-event detection, episode reconstruction, and longitudinal reasoning. The authors argue clinical memory differs structurally from conversational memory — histories span multiple systems, frequencies, and years. Evaluated approaches include vector memory (mem0), temporal knowledge graphs (Graphiti), and agent-managed memory (Letta).
More from Research
- One picture each: gradient descent and backprop explained without arithmetic — alfcnz · 2026-09-24
- Neural net series: multi-layer nets and why cross-entropy and squared error share the same gradient — alfcnz · 2026-09-24
- Teaching Thread: From Perceptrons to Hidden-Layer Representations in Driving — alfcnz · 2026-09-24
- Simulated fly on LSD: connectome wired with eyes shows T4/T5 activity up 10–20% — Merzmensch · 2026-09-24
- Anthropic's first biology lab result draws scrutiny: only ~4 parallel agents and no wet-lab details — basedjensen · 2026-09-24
- Predict rates, not yield: ML kinetic maps explain nickel cross-coupling failures — bravo_abad · 2026-09-24