Decision Context: Persistent Behavioral Learning for AI Agents
Summary
- LLM agents traditionally fail to learn from corrections across sessions, requiring users to repeatedly fix identical mistakes.
- The authors introduce Forge Decision Context (forge_dc), a system that captures behavioral signals from corrections and tool-call failures.
- Forge_dc compiles behavioral signals into imperative RULE directives and injects them precisely at relevant decision points.
- Evaluated on DC-Bench across 71 scenarios, forge_dc achieves 95% compliance compared to 55% with naive retrieval while using 10x fewer tokens.
- The study demonstrates that explicit directive representation, precise activation, and contextual placement are central to persistent behavioral learning.
Why it matters
This research provides a more efficient and effective method for LLM agents to retain procedural corrections across sessions, significantly improving reliability and reducing token overhead for developers.