Agentic Memory Management Is the Next Frontier of Information Retrieval

Abstract: Information retrieval has long been focusing on connecting people with information they need to complete their tasks. Recently, the meaning of this has changed dramatically as more and more AI agents now consume, retain, and act on information on a user's behalf. Consequently, information management now also takes place \emph{inside} the agent in the form of memory management. In this perspective paper, we argue that \emph{agentic memory management} (AMM), which studies the structure, analysis, organization, storage, searching, retrieval, and \emph{use} of information in an intelligent agent's memory, is a natural extension of IR research and an area in which the IR community can make distinctive contributions. Specifically, we first present a system-level framework for AMM and introduce a computational taxonomy of memory objects as \emph{constants}, \emph{functions}, and \emph{operators}, each with distinct retrieval semantics. Then, we compare AMM with established IR research and identify their difference and connections. Last but not least, we outline a research agenda covering benchmarks, candidate principles of memory utility, optimization procedures that mix parametric and non-parametric updates, and more. We hope this paper could attract more people, both inside and outside the IR communities, to think more about AMM problems and research from IR perspectives.

• The paper introduces the concept of agentic memory management (AMM) as an extension of information retrieval, where AMM focuses on how AI agents process, store, and retrieve information to guide their actions.
• AMM is distinguished by its emphasis on the agent's internal memory operations, including learning, forgetting, and utilization of information, which differs from traditional IR by prioritizing tasks relevant to the agent's behavior and performance.
• The paper proposes a research agenda focused on evaluating AMM systems through state quality, utilization, task outcomes, and cost, aiming for robust and effective agent memory systems.

The paper argues that agentic memory management (AMM) should be understood as a natural extension of information retrieval (IR) rather than as an isolated subfield of LLM engineering. Its central observation is that the consumer of retrieved information is changing. Traditional IR generally supplies information to a human who interprets, evaluates, and applies it. In agentic systems, an AI agent may instead read the information, retain selected portions, use them to determine subsequent actions, and modify its behavior over multiple tasks. Consequently, the core IR problems of representation, organization, retrieval, and evaluation now arise inside the agent itself.

The paper is a perspective and framework contribution rather than an empirical study. It does not report a new AMM algorithm or a controlled benchmark result. Its primary contributions are conceptual: a system-level decomposition of AMM, a computational taxonomy of memory contents, an analysis of how AMM differs from established IR, and a research agenda centered on evaluation, memory utility, hybrid optimization, and human oversight (2610.32464).

From conventional IR to agentic memory management

The paper begins from Salton’s classical definition of IR as concerned with the “structure, analysis, organization, storage, searching, and retrieval of information.” This definition applies across Boolean retrieval, learning-to-rank, recommender systems, question answering, and conversational assistants. Historically, however, IR has generally treated the human as the ultimate consumer of retrieved information. User behavior, cognitive constraints, relevance judgments, interface interaction, and task completion have therefore shaped both system design and evaluation.

Generative AI weakens this division between retrieval and downstream use. Search systems increasingly summarize and synthesize documents before presenting information to a person, while agents may execute multi-step tasks such as booking travel, modifying software, or preparing literature reviews. In these settings, retrieved information becomes an internal state that influences planning and action. The agent must decide what to acquire, what to store, what to discard, when to retrieve, and how retrieved information should affect execution.

The paper defines AMM as the study of the “structure, analysis, organization, storage, searching, retrieval, and use of information in the memory of an intelligent agent.” This definition is intentionally broader than vector-store retrieval. It includes conversational state, user facts, episodic records, executable skills, preferences, evolving knowledge structures, and potentially updates to model parameters. The paper’s claim is not that existing IR techniques transfer unchanged, but that IR provides the appropriate intellectual and methodological foundation for studying these operations.

The transition is motivated by several observations about agent consumers. In retrieval-augmented question answering, highly ranked passages that do not contain the answer can reduce accuracy, while randomly selected passages have reportedly improved accuracy by as much as 35% in some configurations [cuconasu2024noise]. This result is explicitly inconsistent with the assumption that retrieval quality can be adequately characterized by conventional relevance scores. Agents also remain vulnerable to indirect prompt injection, because retrieved content can contain instructions that are interpreted as commands rather than merely as evidence [greshake2023indirect]. These effects imply that agent-oriented retrieval must model not only whether an item is relevant, but also how the agent interprets, prioritizes, and operationalizes it.

The presentation problem also changes. Human search interfaces exploit sequential attention, rank position, snippets, and visual scanning. LLMs exhibit a different utilization profile: information placed near the beginning or end of a long context is often used more effectively than information placed in the middle [liu2024lost]. Therefore, conventional rank-discounted metrics such as DCG may be poor proxies for agent utility. Selection, ordering, compression, attribution, and structured formatting become first-class retrieval variables. RECOMP illustrates this shift by compressing retrieved documents to improve downstream model performance rather than human readability [xu2024recomp].

AMM as an internal IR system

The paper distinguishes two changes that are often conflated. The first concerns the consumer: an agent replaces or mediates the human consumer of retrieved information. The second concerns the position of IR: retrieval may remain an external tool, or it may become an internal module that directly shapes the agent’s state and behavior.

External agent-oriented retrieval is represented by systems such as WebGPT, ReAct, Toolformer, and Search-R1. These systems enable an agent to issue searches, invoke tools, and incorporate external evidence into reasoning. The retrieval system remains conceptually separate from the agent, although the agent may learn when to call it and how to formulate requests.

AMM concerns the stronger internal position. A memory system determines what is written after an episode, what is retained or forgotten, what is retrieved for a later task, and how the retrieved material conditions the agent’s actions. The paper uses existing systems to illustrate different forms of internal memory:

• Generative Agents use memories and synthesized reflections to influence planning and interaction [park2023generative].
• Reflexion stores verbal feedback from failed attempts and uses it in later attempts [shinn2023reflexion].
• Voyager stores executable skills acquired in an environment and retrieves them for subsequent tasks [wang2023voyager].
• MemGPT and related systems manage persistent conversational state and user-specific information [packer2023memgpt].
• Mem0 and A-MEM emphasize extraction, consolidation, and organization of long-term information [chhikara2025mem0; xu2025amem].

The implication is that memory is currently one of the most practical mechanisms for post-deployment adaptation. Updating model parameters after every user correction or task is generally infeasible, so deployed agents often improve only through the information they store and later retrieve. This makes memory management part of the agent’s learning process rather than a passive storage layer.

A computational taxonomy of memory objects

The paper’s principal conceptual contribution is a taxonomy that organizes memory into constants, functions, and operators. The taxonomy is intended to complement psychological distinctions such as episodic, semantic, and procedural memory by specifying how different objects should be matched, composed, updated, and evaluated.

Memory object	Content	Retrieval criterion	Typical failure
Constants $C$	Facts, events, user attributes, outcomes	Content and temporal validity	Staleness, contradiction, fabricated detail
Functions $F$	Skills, workflows, reusable procedures	Applicability to the current task	Over-specificity, non-transferability, execution failure
Operators $H$	Preferences, values, styles, behavioral constraints	Scope and applicability	Over-generalization, conflict, silent nonuse

Constants are discrete information items: a passport expiration date, an API endpoint, or the result of a previous attempt. Their retrieval resembles conventional fact retrieval, although temporal validity and contradiction are especially important.

Functions represent mappings from inputs to outputs. A workflow for changing an airline seat or repairing a particular build failure is not useful merely because it is topically similar to the current request. It must be applicable to the current task and environment. This introduces a retrieval criterion closer to procedural transfer than to document relevance.

Operators are higher-order objects that modify how functions are executed. A preference such as avoiding overnight flights or minimizing code changes applies across multiple tasks. Operators are therefore retrieved by scope: the relevant user, context, task class, and strength of the preference must be established before application. A system that retrieves a correct fact but applies an inappropriate operator can still produce an incorrect action.

The paper represents the memory state as $M=(C,F,H)$ and characterizes behavior as the application of an operator-modified function to an input using relevant constants. Although the notation is not presented as a computational model, it exposes three important properties. First, retrieval occurs at multiple levels with different matching semantics. Second, errors compose across levels: correct facts cannot compensate for an incorrectly selected skill or an overgeneralized preference. Third, updates have different scopes. Correcting a fact may affect only executions that use it, whereas correcting an operator may alter behavior across many functions.

The taxonomy also provides a theory of abstraction. Repeated episodes can be stored as constants; recurring patterns can be lifted into reusable functions; recurring corrections across functions can be lifted into operators. The paper emphasizes that these lifting operations remain largely unresolved. Abstraction can improve transfer, but it can also make an item applicable in contexts where it should not be used.

Learning, forgetting, and utilization

The framework groups AMM operations into learning, forgetting, and utilization.

Learning includes adding and updating memory, selecting what to record from an interaction, determining representation and granularity, reconciling new information with existing items, and lifting episodic records into more abstract procedures. These are IR-like write-policy problems, but the collection is interaction-derived, dynamic, personalized, and often small relative to the web.

Forgetting includes deletion, decay, merging, compression, and consolidation. It is not merely a storage optimization. Forgetting can reduce interference, control retrieval cost, address stale information, and satisfy user deletion requests. However, estimating the future value of an item is difficult. Recency, frequency, importance, and model-based judgments are only heuristics, and the value of an unused memory may not be observable at the time it is deleted.

Utilization consists of searching memory and allowing retrieved items to alter behavior. Unlike traditional search, the query may not be explicitly formulated by a person. It must be constructed from the agent’s current state, task, history, and anticipated action. Utilization therefore includes both candidate generation and downstream interpretation. The system must determine not only whether an item is retrieved but whether it enters context, is used correctly, and changes the selected action.

This distinction motivates an agent analogue of click models. In web search, clicks provide noisy evidence about user utility and position bias. In AMM, an equivalent utilization model would predict whether a memory item, given its position, format, provenance, and task context, changes agent behavior. Such a model could support diagnosis, offline evaluation, and optimization. The paper does not provide such a model; it identifies it as a central research problem.

Evaluation and the observability inversion

The paper’s strongest methodological claim is that AMM creates an “observability inversion.” Conventional IR seeks to improve task completion but often observes only queries, rankings, clicks, and partial interactions. Because downstream outcomes such as completing a trip booking or fixing a bug are unavailable, IR relies heavily on proxies such as relevance, ranking quality, and user interaction.

AMM potentially makes downstream use more observable. Researchers can inspect which memory items were retrieved, whether they entered the context, whether the agent’s behavior changed, and whether the task succeeded. The same backbone and task can be rerun with different memory states, enabling controlled counterfactual comparisons that are usually difficult with human users. This creates an opportunity to connect memory decisions more directly to outcomes.

The claim has important qualifications. End-to-end success conflates memory quality with backbone competence: a capable model may compensate for a poor memory, while a weak model may fail despite a useful one. Conversely, outcome-only evaluation does not identify whether failure arose during writing, retention, retrieval, or use. The paper therefore proposes a layered evaluation stack:

1. State quality: whether the right information is stored at the correct abstraction level without contradiction.
2. Utilization quality: whether the information is retrieved when needed and changes behavior appropriately.
3. Task outcome: whether the agent completes the task under a specified backbone.
4. Cost: token usage, latency, storage, and related operational resources.

The paper argues that component metrics should not be discarded. Their role should change from stand-alone proxies for an unobservable objective to diagnostic measurements explaining an observable outcome. This is a significant departure from conventional test-collection evaluation.

The paper also highlights several empirical concerns in existing AMM benchmarks. Current datasets frequently emphasize conversational fact recall, while learning and forgetting receive less attention. Recent analyses report saturation, questionable automated judging, backbone sensitivity, and incomplete cost reporting [jiang2026anatomy]. The paper consequently recommends evaluations that separate constants, functions, and operators; test learning and forgetting as well as utilization; include streams of corrections and delayed outcomes; evaluate scope leakage across users and sessions; replicate results across at least two backbones; and report token, latency, and storage costs.

A research agenda grounded in IR

The paper proposes four primary directions.

Systematic evaluation protocols

AMM benchmarks should follow the logic of Cranfield, TREC, and NTCIR while expanding beyond fact recall. They should include misleading, stale, contradictory, and deletion-sensitive episodes. Multi-user and multi-session settings are necessary to test whether a preference learned in one context leaks into another. Static memory stores are insufficient for evaluating continual improvement because optimization requires feedback streams and delayed consequences.

The proposed benchmark design also treats cost as part of system quality. An accuracy improvement obtained through substantially larger contexts or repeated retrieval may not be preferable once latency and token expenditure are included. This follows established IR practice in which effectiveness, efficiency, and resource constraints are evaluated jointly.

A theory of memory utility

The paper introduces three conjectures rather than established principles.

The Memory Utility Principle states that retention and retrieval should prioritize expected marginal contribution to future task reward, net of storage, retrieval, context occupancy, and interference costs. This generalizes the intuition behind the probability ranking principle, but the paper acknowledges two major difficulties: the future task distribution is unknown, and item utility is non-additive.

The abstraction trade-off conjecture predicts that lifting constants into functions and functions into operators increases transfer while increasing misapplication risk. Aggressive abstraction should therefore help on repetitive task distributions and hurt on heterogeneous distributions. This is a testable claim about the relationship between generalization and contextual specificity.

The minimal scope conjecture states that corrections should initially be applied at the narrowest scope that explains them: session before user, user before cohort, and cohort before global. Scope should widen only after repeated independent evidence. This directly targets personalization failures caused by overgeneralization.

The paper is appropriately explicit that these are hypotheses requiring empirical validation. Their value lies in converting vague analogies to human memory into testable system-level claims.

Hybrid parametric and non-parametric optimization

AMM systems can improve through parametric updates to model weights and non-parametric updates to memory content and organization. The paper argues that current systems lack a principled analogue of gradient descent for textual memory updates. Asking an LLM to revise notes can work in practice, but it is unclear when repeated revision converges, whether it produces monotonic improvement, or how rephrasing can be distinguished from substantive correction [yuksekgonul2024textgrad].

The paper identifies four technical problems: characterizing convergence of non-parametric updates, assigning delayed outcomes to memory decisions, jointly deciding what belongs in parameters versus external memory, and optimizing memory structure rather than only memory text. These problems connect AMM to counterfactual learning-to-rank, online experimentation, continual learning, and structured data management, but the paper does not claim that existing solutions are directly transferable.

Agent analysis and human control

Agent interaction logs create an analogue of query-log analysis. Researchers can study when agents retrieve, how often retrieval is unhelpful, which memory operations precede failure, and how information is formatted and ordered. This could support taxonomies of agent information needs and utilization models analogous to web-search taxonomies and click models.

The paper simultaneously emphasizes that AMM cannot be treated as an exclusively autonomous optimization problem. Persistent personal memory requires inspection, correction, and deletion interfaces. Deletion must remain meaningful even when content has been abstracted into a skill or consolidated into parameters. Memory also introduces attack surfaces, including poisoning and membership inference [chen2024agentpoison; chen2026membership]. These are security and governance issues, but they also concern retrieval, scope, provenance, and state management.

Limitations and open questions

The paper’s principal limitation is that it establishes a research program rather than validating its framework experimentally. No new benchmark, utilization model, memory policy, or optimization algorithm is evaluated. The proposed taxonomy may clarify design space, but its empirical adequacy remains untested. In particular, it is unclear whether constants, functions, and operators are sufficiently expressive for all important memory objects or whether real systems will require overlapping, hierarchical, or probabilistic representations.

The observability inversion also depends on strong assumptions. Agent behavior is measurable in principle, but behavioral traces do not automatically reveal causal memory use. An item may enter context without affecting the action, or an action may appear successful because the backbone already possessed the relevant knowledge. Repeated execution improves control but does not eliminate non-stationarity across model versions, prompting strategies, tool environments, and sampling configurations.

The proposed utility principle leaves unresolved how future task distributions, interference costs, and marginal contributions should be estimated. Minimal-scope updates may reduce leakage but could also delay useful generalization when evidence is sparse. Similarly, the abstraction trade-off predicts a task-distribution-dependent optimum without specifying how a system should infer that distribution online. Finally, the paper leaves open how deletion guarantees can be enforced after information has been lifted into functions, operators, or parameters.

These limitations are not incidental; they define the technical core of AMM. The specific questions are whether utilization can be modeled causally, whether memory utility can be estimated before use, whether non-parametric updates can be given convergence or safety guarantees, and how deletion and provenance can survive abstraction and consolidation.

Conclusion

The paper presents AMM as an extension of IR in which the retrieved information is consumed by an agent and can directly modify its future behavior. Its central framework distinguishes memory objects—constants, functions, and operators—from the procedures that learn, forget, and utilize them, and from the optimization loop that assigns feedback and updates the system. This decomposition explains why conventional content-based retrieval is insufficient for skills, preferences, temporal facts, and agent-mediated action.

The paper’s most consequential methodological proposal is to evaluate memory systems through a combination of state quality, utilization, downstream task outcomes, and cost. Its research agenda calls for benchmark protocols that test learning and forgetting, theories of memory utility, hybrid parametric and non-parametric optimization, and persistent human control. The framework is intentionally provisional, but it provides a technically grounded basis for treating agent memory as an IR problem rather than merely as prompt engineering or storage infrastructure.