How to Use Memory in Agent Builder

Izzo You know that moment when you're explaining the same context to ChatGPT for the fifth time this week?

Izzo Welcome back to Exploring Next, episode two-oh-six. I'm here with Boone, and today we're diving into LangChain's approach to solving that exact problem with Agent Builder.

Boone Right, and this isn't just another chatbot memory gimmick. They've shipped something that actually learns from your corrections and gets better over time.

Izzo Exactly. This matters because we're hitting the limits of stateless interactions. Every conversation starting from zero is... honestly, it's broken UX.

Boone So let's talk architecture. Agent Builder sits on top of Deep Agents, which is LangChain's open source framework for long-running autonomous tasks.

Izzo Boone, break down what that actually means in practice.

Boone Think of it like this — your agent gets an LLM for reasoning, tools like web search and Slack integrations, the ability to spawn subagents for complex tasks, and most importantly, a persistent filesystem.

Izzo And that filesystem is where the magic happens?

Boone Exactly. They've got two types of memory. Short-term lives in the current thread — plans, tool outputs, task progress. Long-term gets saved to a persistent /memories/ path as standard Markdown files.

Izzo Wait, Markdown files? That's surprisingly... simple.

Boone That's the brilliant part. No proprietary database, no complex schemas. Just files the agent can read and write. Your core instructions and skills all live there.

Izzo Okay, but here's my product manager question — how does the user actually make this work? Like, what's the interaction model?

Boone Three main approaches. First, you explicitly tell it to remember. 'That approach worked well, update your instructions to always use that going forward.'

Izzo So it's conversational memory management.

Boone Right, but it gets smarter. If you give clear feedback like 'change your writing style to be more direct,' it'll auto-propose updating its instructions and ask for approval.

Izzo I'm giving that a solid A-minus for user experience. What about the skills system you mentioned?

Boone This is where it gets really clever. Skills are contextually-loaded memory. Think reference library instead of trying to memorize everything upfront.

Izzo Why does that matter?

Boone Because more context isn't always better. An agent holding onto everything can lose focus and start hallucinating. Skills let you say 'create a skill for LangSmith that includes these product features' and it only loads when relevant.

Izzo So if I'm writing about three different products, I'd have three separate skills, and it pulls the right one based on context?

Boone Exactly. Core instructions handle your voice and style, skills handle specialized knowledge. Much cleaner separation of concerns.

Izzo And the third approach is direct editing?

Boone Yeah, you can just edit the Markdown files directly. It's faster for small tweaks, plus you get to see exactly how your agent thinks — like reviewing a teammate's project plan.

Izzo That's actually huge for debugging. Most AI tools are complete black boxes.

Boone Right, and if you spot a wrong assumption or unnecessary step, you fix it in seconds instead of trying to prompt your way around it.

Izzo This feels like a real shift toward AI tools that actually integrate into workflows instead of being one-off interactions.

Boone Totally. And since it's built on Deep Agents, which is open source, you're not locked into their specific implementation.

Izzo Alright, so what should people actually go build with this? First, install Agent Builder and set up a simple agent for something you do repeatedly. Email responses, code reviews, whatever. Then explicitly train it on your preferences. Second, experiment with the skills system. Create a skill for your company's specific context or a technical domain you work in regularly. And definitely peek under the hood — browse those memory files to understand how it's reasoning. Adding th