Research notes on agent memory, trustworthy multimodal AI, and medical AI evaluation.
SYNAPSE, accepted to Findings of ACL 2026, turns long-term agent memory from semantic search into spreading activation over an episodic-semantic graph — new state of the art on LoCoMo with 95% fewer tokens per query.
MedVIGIL is a clinician-supervised benchmark for testing whether medical vision-language models fail safely when the visual evidence contract is broken.