decision-context

人間とAIが現実との対話を継続するために、認知活動を理解し、その認知活動を継続可能にする方法を研究するプロジェクト。

Research on cognition and its operation to sustain dialogue with reality.

DC-Bench: a benchmark for behavioral memory — does an agent apply a past correction to an unseen task, not just recall it? 71 scenarios across 9 dimensions, built from real agent traces, with control vs. treatment runs and LLM-judged lift. Approach- and provider-agnostic.

Behavioral memory for LLM agents: compile corrections and tool failures into scoped RULE directives injected at the decision point. Five self-contained stages, 55% → 95% on DC-Bench.