# Giving AI Agents Memory That Learns — Jake Broekhuizen, LangChain **URL:** \ **Category:** Content Creators\ **Tags:** self-adapting, agents, ai-engineer\ **Created:** [8 October 2026 14:00 UTC](https://www.artofsm.art/t/giving-ai-agents-memory-that-learns-jake-broekhuizen-langchain/25275 "2026-10-08T14:00:37Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [8 October 2026 14:00 UTC](https://www.artofsm.art/t/giving-ai-agents-memory-that-learns-jake-broekhuizen-langchain/25275/1 "2026-10-08T14:00:37Z")
[![](https://www.artofsm.art/uploads/default/original/3X/2/3/23eb0a683b92186f39981dcf2e28f4b1cb6d82cc.jpeg "Giving AI Agents Memory That Learns — Jake Broekhuizen, LangChain") ](https://www.youtube.com/watch?v=KGFyOtl5ktI) Jake Broekhuizen from LangChain discusses enabling AI agents to learn and improve over time by converting interaction traces into durable long-term memory, focusing on procedural memory to refine agent behavior such as tone compliance in financial advice. He presents LangChain’s memory loop implementation via LangSmith and the context hub, which captures, analyzes, and updates agent memory to create a continuous feedback system that enhances reliability while emphasizing selective memory updates and human oversight. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [8 October 2026 14:24 UTC](https://www.artofsm.art/t/giving-ai-agents-memory-that-learns-jake-broekhuizen-langchain/25275/3 "2026-10-08T14:24:41Z")
Jake Broekhuizen from LangChain presents on the topic of giving AI agents memory that learns, emphasizing the importance of enabling agents to improve from one interaction to the next. He illustrates this with a financial services agent designed to provide spending and budgeting advice while adhering to strict tone guidelines. The challenge faced was that the agent’s tone sometimes slipped into being too directive, which violated compliance rules. Correcting these issues manually was not scalable, highlighting the need for a system where the agent can continuously learn and adapt based on its experiences. Broekhuizen explains the distinction between traces and memory in AI agents. Traces are logs or transcripts of what happened during an agent’s run, capturing tool calls, reasoning, and artifacts. However, these traces only become memory when the lessons learned are converted into durable context that the agent can reference in future runs. He introduces a taxonomy of memory inspired by cognitive science: semantic memory (facts and preferences), episodic memory (past experiences and patterns), and procedural memory (instructions and skills). Procedural memory, which governs how the agent behaves, is particularly crucial for driving visible improvements. Another key concept discussed is the difference between working (short-term) memory and long-term memory in agents. Working memory holds temporary context needed for the current task, such as intermediate results or retrieved files, while long-term memory stores durable information like skills and rules for future use. The agent’s operation involves a read-write cycle where relevant long-term memory is loaded into working memory at the start of a run, and after the run, useful signals from the trace are filtered and written back into long-term memory. This cycle enables agents to adapt and improve over time. Broekhuizen describes LangChain’s implementation of this memory loop through LangSmith and the context hub. LangSmith provides observability by capturing detailed traces of agent runs, while the context hub performs background analysis to extract meaningful signals from these traces. These signals are then used to update the agent’s long-term memory, such as modifying procedural rules to correct tone issues. This approach creates a feedback loop where the agent’s behavior is continuously refined based on real-world interactions, making it more reliable and effective. Finally, Broekhuizen highlights important considerations when designing memory systems for agents. Not all trace data should become memory; only relevant signals that improve the agent’s behavior should be promoted. He also warns about challenges with caching and ensuring memory updates are accessible in long-running agents. Additionally, he stresses the importance of human oversight for critical procedural memory updates, such as tone and policy changes, to prevent unintended consequences. Overall, memory is framed as the mechanism by which experience becomes context, enabling agents to learn and perform better with each run. ## Useful Links - [LangChain Labs Official Blog](https://www.langchain.com/blog/introducing-langchain-labs) — Directly explains the LangChain Labs tools and concepts for agent memory and learning.