How should AI agents organize their memory for users?

A CHI 2025 Extended Abstracts study by Brennan Jones, Kelsey Stemmler, Emily Su, Young-Ho Kim and Anastasia Kuzminykh — the authors call it "a preliminary investigation" — interviewed six regular users of personalized AI tools with long-term memory and thematically analyzed 54 Reddit discussion threads about memory features. They report that "users generally consider agents' memories as belonging to different categories along a hierarchy from more generalized knowledge to more specific knowledge about the user or task." At one end sits "generalized knowledge (e.g., factual memories of general knowledge facts, usually contained in the model's training data or fetched externally through retrieval-augmented generation, or RAG)"; in the middle, "knowledge about the user (e.g., their personal or social details, preferences, etc.)"; at the other end, "more specific knowledge about the domains, projects, or tasks that the agent assists the user with." Participants wanted "the system's memories to be cleanly organized by these categories," and the authors note that these needs "evolve depending on the task or stage of the activity they are working on."

The paper's mechanism for the hierarchy is practical, not theoretical: because no tool surfaces these tiers directly, participants built their own approximations — "organizing, separating, and abstracting access to memories across different tasks, projects, and domains, including their use and organization of chat threads, customs prompts and instructions for defining personas and roles, fine-tuning models, and using different user accounts." A new chat thread becomes a crude task-boundary; a custom-instructions block becomes a crude persona-boundary; a second account becomes a crude domain-boundary — each a substitute for a hierarchy the interface itself doesn't expose.

This complements rather than competes with What mental models do users hold about LLM memory?: that study's four models (transient buffer, training-data extension, active processing, don't know) describe what participants believe memory technically does, while this hierarchy describes what kind of content participants want sorted, independent of which mechanism they believe produces it — a user holding any of those four mechanism-beliefs could still want general, personal and task knowledge kept apart. It also sharpens the design mandate in What memory controls do users actually want from LLMs?: where that note's participants asked for consent, editing and deletion controls without specifying what the controls should be organized around, this hierarchy supplies a candidate axis — control surfaces scoped to generalized, personal or task-specific tiers rather than one flat list.

The excerpt doesn't weight the tiers against each other — it doesn't say which one participants cared most about controlling, or how often the hierarchy broke down in practice — and the evidence base is small and preliminary by the authors' own framing: six interviewees (four women, two men, ages 24–35) reimbursed CAD$20, plus an unweighted, self-selected set of 54 Reddit threads surfaced by keyword search. What the evidence does support, at that modest strength, is a concrete design target: memory interfaces exposing a visible general/personal/task tier structure would be meeting an organizing scheme participants were already building for themselves by hand, not inventing one from scratch.