Every LLM agent prototype starts with the same uncomfortable workaround: rebuild the user's context, stuff more into the prompt, retrieve a few documents, and hope the next turn has enough information. That works for demos, but it breaks down when agents need to return to the same user, remember preferences, reuse shared knowledge, and keep state across sessions. This talk shows a practical memory-tool pattern for LLM agents. We will look at where chat history, prompt stuffing, and ad hoc retrieval fall short, then walk through an explicit memory layer where agents can list, read, write, patch, and search durable memory as part of their normal tool loop. The demo will use MemexAI, an open-source Postgres-backed memory layer, but the goal is to leave attendees with a reusable design pattern they can adapt to LangChain, Vercel AI SDK, LlamaIndex, or their own agent stack. Attendees will learn: - How to separate retrieval, chat history, and durable agent memory. - How to design memory tools with scoped user/shared state, revisions, and inspectable writes. - How to inject memory back into the model context without hiding state inside opaque prompts.
Sooraj Sanker is a Bengaluru-based fullstack and AI engineer, and the founder of MemexAI, an open-source Postgres-backed memory layer for long-horizon AI agents. He spent roughly six years at Hasura and PromptQL, where he worked on conversational analytics, AI agent interfaces, tool-based workflows, artifact-aware chat, and state management for distributed agent execution. His work sits at the intersection of agent infrastructure, context engineering, data UX, and production reliability.
