User as Code: Executable Memory for Personalized Agents

A team of researchers has proposed a new architecture for AI agent memory called User as Code, which represents a user's history and preferences as executable Python code instead of unstructured text or knowledge graphs, according to a paper submitted to arXiv in June 2026 [1]. The paper, titled "User as Code: Executable Memory for Personalized Agents," argues that current memory systems for personalized AI agents store user information as a "bag-of-facts" — unstructured text, knowledge graphs, or flat stores — and retrieve it by searching for entries similar to a new request [1]. This approach, the authors contend, recalls individual facts well but struggles with tasks requiring reasoning across many stored records, such as resolving contradictions or answering aggregate questions [1]. User as Code (UaC) instead models the user as a living software project. Typed Python objects hold the user's state, and ordinary Python functions encode the rules that govern it, so that representing and reasoning about the user occur in a single medium an interpreter can run [1]. The system relies on a two-phase pipeline: an append-only log that never discards a fact, which is periodically checkpointed into typed code [1]. The performance gap is stark on aggregate questions. On the LOCOMO long-term conversation benchmark, UaC achieved 78.8% recall, matching both a full-context upper bound and the strongest prior memory systems [1]. But when asked questions that require computation over a user's history — such as "how many international trips did I take last year?" — retrieval-based memory fell to between 6% and 43% recall, while UaC remained near-perfect at 99% [1]. The paper attributes this to the fact that the answer becomes a one-line computation over typed state rather than a search over text [1]. The deterministic execution of rules also enables a capability that query-driven memory cannot provide: unsolicited, safety-critical alerts. The paper gives the example of a newly prescribed drug that conflicts with an allergy recorded months earlier, which UaC can surface automatically whenever the state changes [1]. The paper was posted on arXiv, the open-access repository of electronic preprints that, as of November 2024, receives about 24,000 submissions per month and has surpassed two million articles [9]. The work appears under arXiv's Artificial Intelligence category and is accessible through arXivLabs, a framework launched in 2020 that allows community collaborators to develop and share experimental tools directly on the platform [7]. arXivLabs projects operate under guidelines requiring partners to share arXiv's values of openness, community, excellence, and user data privacy, with third-party collaborators granted access only to minimal and anonymized user data [7].