LLM agents, which are programs that can use tools and make decisions, don't learn from past mistakes between different conversations. Every time a user starts a new chat, the agent forgets what it learned before, so the user has to repeat corrections like "always check security settings first" over and over.
The researchers built a system called Forge Decision Context (forge_dc) that captures both human corrections and the agent's own failed attempts. Unlike memory systems that only store facts, this system turns these signals into specific "rule" instructions that are inserted at the exact moment the agent needs to make a decision.
On a new benchmark of 71 scenarios, forge_dc achieved 95% compliance with past corrections, compared to 55% for simple retrieval. It used about 10 times fewer extra tokens per query, making it both more effective and more efficient for persistent learning.