Talks

Say you are building an AI agent that must provide human-like experiences to its users. To achieve this, you must feed the LLM with the right data. Then it happens: you create a tool to query your SQL database for transactions, another for RAG on factual data, a third for recent conversations, and so forth so on. Every memory data source requires its own retrieval tool, and suddenly, your agent needs a dozen different functions to remember things. What started as a simple agent implementation has devolved into a complex web of tools and databases, where debugging why an agent failed to perform a task requires tracing through multiple systems and tool calls.

The real challenge here isn't just about merging data systems; it's about rethinking how memories are utilized and the data access patterns they require for agents to answer questions efficiently. It's similar to how humans handle memories: we store them in different parts of the brain, yet we can retrieve them without worrying about things like data types, storage approach, or how they relate to each other.

This talk demonstrates how to design and implement a unified memory layer that consolidates all agent data access. You'll learn to architect a single memory abstraction that handles both working memory and persistent knowledge, replacing dozens of retrieval tools with elegant, unified operations. You'll leave with practical knowledge to build agentic AI systems that can provide better user experiences without the need to develop and maintain a data integration nightmare.