AI Agent Memory (Part 1): Why Remembering Isn't the Same as Learning

AI agents retain experience across histories, notes, and reusable instructions, but the retrieval process requires more than just memory. Their models possess broad knowledge gained during training, yet memory must enable the system to discern what matters, recognize situations, and select appropriate actions. Christopher Nolan's Memento illustrates this concept through Leonard's struggle to form lasting memories, relying on photographs, notes, and tattoos.

Language models typically don't modify their parameters after each conversation. Instead, software surrounding the model saves the history and retrieves information when work continues. While the model already possesses extensive knowledge, it can reason about new information and validate its records. This analogy highlights a key question: how does yesterday's note aid in today's decision-making?

Memory encompasses four aspects: working memory, episodic memory, semantic memory, and procedural memory. They work together, not sequentially, and are distinct from one another. An agent's log can assist in utilizing past events, while written instructions can help carry out tasks. However, these files alone do not confirm that the agent remembers or behaves like a human.

During training, a model acquires language, facts, relationships, and problem-solving strategies. This knowledge is dispersed across its parameters and not a collection of pages for the model to search. When used, the model receives context, such as instructions, conversation, search findings, and tool results. Corrections to context can alter the model's response without changing its parameters.

The harness, or software surrounding the model, manages tools, tracks tasks, and determines the model's subsequent input. Saved preferences and tested scripts can enhance future responses, but context is crucial for saving and retrieving useful information.

An experienced cook quickly adjusts heat based on sizzling sounds, while a beginner might miss such warnings. Practice enables the cook to notice useful information. Terms like goal, attention, recognition, belief, and judgment describe this process. Experience can influence attention even without conscious recall, but humans still make errors and forget. Experience can modify how we perceive new situations.

When an agent saves a log, writes a lesson, and stores it, search or a selector later selects what to bring into context. However, several issues can arise. The lesson may omit exceptions, have an inaccurate description for the new task, or the agent might retrieve the correct note but face a changed situation. Determining whether yesterday's experience aids the agent in noticing the right thing and making a better choice today is vital.

Ultimately, the question remains: does the agent spot familiar risks, update outdated expectations, or select a more reliable method with fewer reminders? External memory provides a readable and correctable record, but more records alone do not guarantee competence. The agent's continued performance well when the situation changes, including using outdated methods, is essential. The next question is how the system identifies which saved lesson is relevant.