Agent Memory
Agent memory is the set of mechanisms by which an AI agent retains, organizes, and selectively recalls information across interactions and sessions. It is what allows a stateless language model to behave as a persistent system that remembers user preferences, preserves task state, learns from earlier mistakes, and reuses procedures. Memory has become a core capability of foundation-model-based agents, underpinning long-horizon reasoning, continual adaptation, and effective interaction with complex environments[^c1], and it is the component that turns a text generator into a genuinely adaptive agent[^c2]. By 2026 it is treated as a first-class architectural layer with its own benchmarks, literature, and vendor ecosystem[^c11].
The field organizes memory in several ways. Cognitive-architecture research treats the language model as the core of a larger system whose internal memory divides into a working store for the current situation and long-term stores for facts, experiences, and skills[^c3]. Working memory holds recent perceptual input, active goals, and intermediate reasoning results[^c4]. Long-term memory is further distinguished by content: semantic memory stores facts, episodic memory stores events and trajectories, and procedural memory stores rules and skills. A complementary view separates memory by form — token-level text, parametric weights, or latent representations — and by function — factual, experiential, and working memory.
Operationally, memory is a lifecycle rather than a single store. It can be described as a write–manage–read loop coupled to perception and action[^c5], in which an agent decides what to record, maintains and revises it over time, and retrieves it when relevant. Implementations differ widely. Some systems page information between a bounded context and external storage, treating the context window as constrained main memory[^c6]. Others retrieve from an external index, as retrieval-augmented generation does when it combines parametric and non-parametric memory[^c7]. More structured systems build knowledge graphs and summarize communities of related entities to answer questions that require global understanding[^c8], while hierarchical designs stack raw observations, distilled facts, belief trajectories, and higher-level schemas into a capability hierarchy rather than a single flat store[^c15].
Designs are converging on a division of labour between the task agent and a memory process. Memory work increasingly runs asynchronously, so that creation and retrieval do not consume tokens or add latency to the active task[^c16]. At the same time, the organization of memory is being handed to the agent itself: systems now ask whether LLM agents can organize their own memory using ordinary documents[^c17] and store it in linked, revisable notes rather than fixed-schema records. Retrieval is also moving from one-shot recall at request time to continuous monitoring: existing methods retrieve primarily from the initial request, with limited awareness of ongoing execution and emerging personalization needs[^c18], which has prompted execution-aware memory agents that decide when to stay silent and when to intervene.
The long-context debate has sharpened the design calculus without settling it. Larger windows postpone but do not remove the cost and attention limits of full-context injection, and the median overlap between a query and its stored history can be low enough that most of a larger window remains irrelevant[^c14]. The practical consequence is that long context and dedicated memory are complements: the window manages the current turn, while external memory manages what happened before and what it implies.
Measurement has matured alongside the mechanisms. Public competitions now run memory systems under a shared protocol: the second Agent Memory Challenge published separate standings for text, code, and multimodal memory[^c20], receiving 402 submissions and releasing 292 valid entries after execution and compliance review[^c19]. Academic benchmarks push the same questions to extreme scales and to action. In agentic settings, a frontier model reaches only about 72 percent accuracy on a long-horizon memory benchmark[^c26], and on a GUI benchmark the best memory-free model reaches 16.4 percent success weighted by path length while end-to-end GUI agents fall below five percent[^c27]. The failure is not only recall of dialogue; it is the ability to carry evidence into action across hundreds of steps.
Memory also introduces new failure modes. Persistent state can become stale, poisoned, or over-scoped, and persistence converts transient, single-episode problems into long-lived, system-level ones that single-episode evaluation cannot see[^c9]. Benchmarks have consequently moved from short conversational recall toward multi-session histories, knowledge updates, conflict resolution, and forgetting-aware scoring; commercial assistants and long-context models show substantial accuracy drops on sustained interactions[^c10]. At the same time, deployed memory has begun to be treated as infrastructure that needs health checks, heartbeats, and rollback rather than as a feature that either works or does not[^c12]. Evaluation has broadened accordingly, since performance and efficiency scores alone cannot show whether a memory system honors its responsibilities to privacy, correction, and deletion[^c13]. The subject therefore spans not only storage and retrieval but consolidation, forgetting, retrieval policy, personalization, security, governance, and evaluation.
Two problems have moved to the centre of research. The first is deciding how to trust memory before it is updated: delayed feedback about a stale fact can be confined to that fact or propagated across the store, and fixed local or global policies each fail in some regime, so online conflict-governance routers that arbitrate between the two scopes outperform the better fixed policy by five to seven points[^c22]. The second is identity and evidence. Once recall saturates, the residual difficulty is deciding which entity a query names, which time-state is current, and when to answer at all[^c28]. Auditable designs tie every stored fact to the exact words that support it[^c24], and multi-agent systems reveal that evidence can be retrieved yet fail to reach the agent that must reconcile it: in one controlled study, workers surfaced both sides of a planted conflict for only 35 percent of the pairs they collectively retrieved[^c25].
Infrastructure has followed the same trajectory. Managed platforms now retain short-term memory events for up to ten years for compliance and returning users[^c23], and combine memory with permission-filtered document retrieval in a single governed request, while leaving authentication, retention policy, and monitoring with the application[^c29]. The direction of travel is consistent: memory is being treated less as a cache of conversation and more as governed persistent state, with its own storage tiers, lifecycle rules, audit trails, and evaluation — a system whose value depends as much on what it forgets and refuses to disclose as on what it remembers.