Giving AI Agents Memory That Learns — Jake Broekhuizen, LangChain
LangChain's lab team argues agents must continuously learn from run traces, not require manual log review
“observing and turning those observations and agent experiences into lessons for future launches are obviously very different things”
Jake Broekhuizen of LangChain's lab team argues that the critical unsolved problem in agent development is converting run traces and observations into durable memory that improves future behavior. He illustrates the cost of the status quo with a fintech case study where an agent's tone drifted into non-compliant territory, requiring expensive manual log analysis and context patching to fix. The talk frames agent memory not as recall of past conversations but as a self-improving feedback loop — a framing that has direct implications for how teams instrument and operate production agents.