What should an agent remember? Agent memory: personalization, self-improvement and organizational memory

By Turing Post · Sat, 17 Oct 2026 15:06:21 +0000.

The Turing Post article grew out of preparations for "Expert Showcase: Agent Memory", an O'Reilly event planned with Nicole Butterfield, executive editor at O'Reilly. The organizers wanted to approach agent memory from several angles: an agent that learns from experience, a robot that interacts with the physical world, and an organization whose knowledge still lives in people's heads (a recurring problem Turing Post already addressed in "The Org Age of AI").

The article recalls that when Turing Post first covered agent memory, it used the film "Memento" as an analogy: trying to make sense of the world with notes and photographs because you cannot form new memories. That image reflects a common frustration with AI. But preparing the event led to a broader question: once an agent can remember, what does that unlock for us, and what becomes possible? The piece proposes a structured way of looking at the topic alongside the people building it, stressing that nothing is solved or settled, and that this is precisely what makes it interesting.

The episode follows on from the previous one in the "AI Builds AI" series, devoted to skills: reusable procedures that help agents do their work. Memory goes one step further back: where those procedures can come from, for example when an agent is corrected, tries something that fails or discovers a better way to complete a task. What is explored is how that experience becomes meaningful and what it takes to keep it useful when circumstances change.

The episode's agenda includes these questions: what is an AI agent's memory; how to turn a record of what happened into something the agent can learn from; how much of the past should enter the present; whether an agent can learn across tasks; what changes when experience comes from the physical world; how to build organizational memory when much of the knowledge is in people's heads; what an agent should remember when no one is sure; how to tell whether memory is helping and what to test first; and what it would take for an agent to question what it thinks it knows.

Defining agent memory: in simple terms, memory gives the agent a way to use information from earlier work in a later decision. That information might be a preference, a fact about a project, a record of a failed attempt or a procedure that worked well. For it to be useful across separate sessions, the system surrounding the agent must retain it and make it available when it is needed again.

The article notes that the term "memory" is confusing because when humans remember something, we rarely think about which part of the process stored it, found it or connected it to the present situation. With agents, someone has to design each of those steps. A larger context window can hold more material during a task, but it does not decide what deserves to survive afterward or how a future agent should use it.

Lewis Liu, who works on enterprise AI at Microsoft, proposes separating three ambitions: personalization, self-improvement and organizational memory. They often come up in the same conversation, yet each demands something different from the system. Personalization helps the agent adapt to a specific person: how they like their reports, which projects they work on, what they have already explained. Self-improvement asks whether experience can change how the agent performs a task. Organizational memory extends the question beyond a single user or agent: can knowledge accumulated within a company be made available to others who need it? Slide 5 of Liu's talk, "The State of Agent Memory", notes that each involves different implementation decisions and adoption challenges.

The article offers a table for recognizing whether each ambition is working. For personalization, the example is remembering the preferred format for a weekly update, and success is shown if the preference carries over to a new session. For self-improvement, the example is learning to check a project's alternative names, and success is the agent finding updates it previously missed. For organizational memory, the example is sharing an approved procedure across teams, and success is other teams applying it correctly and being able to trace or correct it.

The same interaction could contribute to all three, but succeeding at one does not prove success at the others. Remembering a preferred format says very little about whether an agent can learn a reliable method, and even less about whether a company should adopt it. By making these ambitions explicit, the useful question becomes: what should be different the next time the agent does its job?

The email cuts off at the section "How do we turn a record of what happened into something an agent can learn from?", because the rest of the article is reserved for Turing Post Premium subscribers. As a result, this summary does not cover the content of the later sections (past context, learning across tasks, memory from the physical world, organizational memory, uncertain beliefs, evaluating memory and questioning one's own knowledge), of which only the questions themselves are available.