Memory Is a Context-Management Problem

AI agents are becoming less like chat interfaces and more like persistent workers. They browse, call tools, modify files, query databases, and continue tasks across multiple sessions. That shift makes memory unavoidable: an agent that cannot preserve the right constraints, discoveries, and decisions is not merely forgetful—it is unreliable.

Recent systems have moved beyond the simple “store a summary in a vector database” pattern. Anthropic describes compaction, structured note-taking, just-in-time retrieval, and sub-agent architectures as different ways to manage a finite attention budget. OpenAI’s newer systems treat memory synthesis, tool history, filesystem state, institutional knowledge, and runtime inspection as distinct layers. Letta pushes the stronger claim that memory is not a detachable feature at all, but a responsibility of the agent harness.

The central difficulty is that remembering more is not necessarily better. Irrelevant history can degrade performance; summaries can erase details that later become important; persistent memories can become stale, contradictory, or poisoned; and personalization raises difficult questions about ownership, privacy, and correction. The system must decide not only what to store, but what to expose, when to update it, and how to know whether it helped.

This seminar follows memory from context design to long-term operation: what counts as memory, how memories are formed and revised, how agents navigate large stores and tool histories, how long-running systems preserve state, and how we evaluate whether continuity improves rather than quietly damages behavior.