How we built Agent Builder’s memory system

LangSmith launched Agent Builder, a no-code platform for creating task-specific agents with memory as a core feature. Unlike general-purpose AI tools, these agents perform the same tasks repeatedly, making memory crucial for learning from previous sessions and avoiding repetitive user instructions. The team built the memory system using a virtual filesystem stored in Postgres, representing agent knowledge through standard files like AGENTS.md for core instructions and custom tools.json for specialized functions, allowing agents to automatically update their capabilities through user feedback and experience.