PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Memory Management Is the Next Frontier of Information Retrieval

Qingyao Ai

arXiv:2610.32464v1cs.IR

TL;DR

As AI agents increasingly consume and act on information for users, information management also takes place inside agents as memory management. The paper presents AMM as an IR problem, develops a framework and taxonomy, and concludes that IR can help build stronger foundations for the area.

  • Problem

    AI agents’ information decisions now occur inside their memory, but the research area needs stronger foundations and systematic evaluation.

    Source: §1 · p. 1, §8 · p. 10

  • Method

    The paper presents a system-level AMM framework, a taxonomy of memory objects, and a research agenda for studying the area.

    Source: §1 · p. 2, §4.2 · p. 5

  • Results

    AMM can be viewed as a new IR problem, and IR can help build stronger foundations for it.

    Source: §8 · p. 10

  • Takeaways & Limitations

    The proposed research directions include systematic evaluation, testable memory-utility principles, hybrid parametric and non-parametric updates, and human-in-the-loop systems.

    Source: §8 · p. 10

    End-to-end task success confounds memory quality with agent competence, while outcome-only evaluation does not diagnose which memory stage failed.

    Source: §5.3 · p. 8

Abstract

from arXiv · show

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.

1 Introduction

As AI agents increasingly process and act on information for users, information management has moved inside agents into memory. The paper frames agentic memory management as an extension of IR and calls for foundational research on it.

1 Introduction

  • AI agents’ internal information decisions—what to collect, retain, discard, retrieve, and use—can affect problem-solving capability and improvement with experience.The paper names this internal information-management problem agentic memory management (AMM).Source: §1 · p. 1
  • AMM studies the structure, analysis, organization, storage, searching, retrieval, and use of information in an intelligent agent’s memory.Source: §1 · p. 1
  • The paper positions AMM as a new IR problem, connecting it to existing retrieval work while raising challenges as IR shifts from serving people to serving agents.Existing AMM studies include conversational state, reusable skills, user facts, and evolving knowledge networks.Source: §1 · pp. 1–2
  • AMM research has diverse terminology and frameworks, limited theoretical grounding, and benchmarks often confined to narrow tasks without well-defined protocols or roadmaps.The paper proposes systematic evaluation, memory-utility principles, hybrid parametric and non-parametric optimization, and human-in-the-loop systems as research directions.Source: §1 · p. 2

2 What Information Retrieval Has Studied

The paper organizes IR research through complementary user and system views, tracing their history and showing how both address the shared goal of helping people satisfy information needs or complete tasks.

2.1 A Brief History

  • IR developed from early indexing and shared Cranfield evaluation through ranking, web search, recommendation, and modern neural methods, while retaining a recognizable research agenda.The agenda analyzes information, matches it to a need, serves it to a user, and evaluates the outcome.Source: §2.1 · pp. 2–3

2.2 The User View

  • The user view studies user behavior, designs services around human cognition, and evaluates systems using assumptions about how people interact with results.Examples include query and click analysis, search interfaces, discounted cumulative gain, diversity, and fairness measures.Source: §2.2 · p. 3

2.3 The System View

  • The system view studies how information is represented, matched, delivered efficiently, and evaluated for quality and cost.Its topics include collection analysis, retrieval and recommendation models, indexing, efficiency, latency, throughput, and resource constraints.Source: §2.3 · p. 3

2.4 Two Views of the Same Problems

  • User and system views approach common problems from opposite directions, and their methods connect through relevance, clicks, ambiguous queries, and evaluation.Together, they serve the goal of helping people satisfy information needs or complete tasks.Source: §2.4 · p. 3

3 From Serving People to Serving Agents

As models and agents increasingly process retrieved information before people see it, IR must reconsider both how agents consume information and where retrieval operates. The paper treats these as parallel changes, with agentic memory management arising when both apply.

3 From Serving People to Serving Agents

  • Retrieved information increasingly passes through a model before reaching people, and IR can either remain an external tool or become an internal module shaping agent behavior.Source: §3 · p. 4

3.1 Consumer Change: From People to Agents

  • Knowledge about human users cannot simply be assumed to transfer when an agent becomes the information consumer.Source: §3.1 · p. 4

3.1.1 Information Processing.

  • Agents process and use information differently from people: irrelevant passages can reduce answer accuracy, random passages sometimes improve it, and retrieved instructions may be followed.Agents can also read many results or issue repeated reformulations, with costs measured in tokens rather than human attention or patience.Source: §3.1.1 · p. 4

3.1.2 Information Presentation.

  • Agent-facing presentation concerns selection, format, granularity, and length, since context position affects accuracy and compressed summaries can target downstream performance.Source: §3.1.2 · p. 4

3.2.1 IR as an External Tool.

  • External retrieval systems return results but do not directly control how agents use them, although agents can adapt to tools through learning when and what to search.Source: §3.2.1 · p. 4

3.2.2 IR as an Internal Module.

  • Internal memory systems decide what agents store and retrieve and how retrieved content conditions action, shaping preferences, plans, experience, and interaction.Memory is described as a practical route to continual learning because retraining for every user, task, and correction is rarely feasible.Source: §3.2.2 · pp. 4–5

3.3 Two Parallel Changes

  • The consumer change and position change are parallel: agent-serving IR must address processing and presentation, while internal memory also raises questions about changing behavior.AMM combines an agent consumer with IR operating inside the agent.Source: §3.3 · p. 5

4 A System-Level Framework for AMM

The framework treats agent memory as writable information outside the current inference call, organized into objects, procedures, applications, and feedback-driven optimization. It distinguishes constants, functions, and operators by how they are retrieved, composed, and updated.

4.1 Definition and Scope

  • Memory is writable information outside the current inference call that can later condition behavior; the framework organizes it into objects, procedures, applications, and optimization.The context window is where memory is used, not memory itself.Source: §4.1 · p. 5

4.2 Objects: Constants, Functions, and Operators

  • Constants match by content, functions by applicability, and operators by scope; the taxonomy also distinguishes their failure modes and update dynamics.A selected skill is modified by an applicable operator and applied using relevant facts.Source: §4.2 · p. 5, Table 2 · p. 6

4.3 Procedures: Learning, Forgetting, Utilization

  • Learning decides what episodes leave behind, forgetting manages interference, retrieval cost, and deletion, and utilization finds memories and lets them shape behavior.Forgetting policies need to estimate future value, while agent-facing formatting and conflict reconciliation remain open questions.Source: §4.3 · p. 6

4.4 Applications: What Memory Is For

  • Four illustrative application families are information access and synthesis, personal assistance, procedural tasks, and long-horizon collaboration; each makes different memory useful.Benchmarks limited to conversational fact recall can favor some application families and neglect others.Source: §4.4 · p. 6

4.5 Optimization: Learning What to Remember

  • Optimization interprets outcomes, corrections, errors, and abandoned interactions at the right scope, then assigns outcomes to memory decisions across a long decision chain.Wrongly scoping a learned preference can cause it to leak from one person into other users’ agents.Source: §4.5 · p. 7

4.6 Existing Systems Are Special Cases

  • Representative systems emphasize constants and utilization, while support for reusable functions, scoped operators, principled forgetting, and learned optimization is less consistent.Table 3 classifies representative systems as central, partial, or absent across framework components and optimization types.Source: §4.6 · p. 7, Table 3 · p. 7

5 Comparing AMM with Established IR

AMM shares IR’s information-management goals and challenges, but differs in how memory is produced, who consumes it, and how its effects can be evaluated. Treating agents as instrumentable consumers makes downstream use more observable, while requiring evaluation to separate memory quality from agent competence.

5.1 Objects: A Difference of Degree

  • AMM and established IR both manage information at scale amid redundancy, staleness, and noise, but memory is often interaction-produced and user-specific.This can mean small logged collections, but also fewer opportunities to aggregate and added obligations for personal data.Source: §5.1 · p. 7, Table 4 · p. 7

5.2 Users: The Agent as a New Kind of User

  • An agent is a model rather than a person, so its attention, use of retrieved content, and propensity to retrieve vary with position, format, and task framing.A proposed utilization model would predict whether a memory item changes behavior and could support diagnosis, offline evaluation, or optimization.Source: §5.2 · p. 7
  • Agents can be repeatedly instrumented and replayed under controlled variations, but memory findings should identify the backbone and ideally replicate across backbones.Memory-system performance varies substantially with the backbone.Source: §5.2 · p. 7

5.3 Evaluation: The Observability Inversion

  • AMM evaluation can observe retrieval, context entry, behavior change, and task success, making downstream use more measurable and topical relevance less informative on its own.The paper calls this shift the observability inversion.Source: §5.3 · p. 8
  • Evaluate memory as a stack of state quality, utilization quality, task outcome under a stated backbone, and cost, because success alone confounds memory with agent competence.Outcome-only evaluation also cannot diagnose whether failure arose in writing, retention, retrieval, or use; cost should accompany quality measures.Source: §5.3 · p. 8

6 Why IR Should Contribute

IR and AMM share the objective of helping complete tasks, and IR methods and methodological experience can inform agent memory systems. The paper argues for IR’s contribution to AMM, not ownership of the area.

6 Why IR Should Contribute

  • AMM preserves IR’s task-completion objective while adding an agent as intermediary, and memory systems resemble small, personal, write-heavy, feedback-rich retrieval systems.Relevant shared methods include indexing, retrieval and ranking, recommendation, clustering, query reformulation, compression, caching, and evaluation design.Source: §6 · p. 8
  • IR brings reproducible user-study methods, simulation, and information-management expertise that can transfer to agent memory research.Its experience with proxy metrics, biased feedback, fair baselines, and shared benchmarks can also inform system and evaluation design.Source: §6 · p. 8
  • The paper frames AMM as an opportunity for IR to shape a transition drawing on its expertise, not as a claim of ownership.The authors compare this opportunity with generative information access, where an IR vision was realized mainly through LLMs developed outside IR.Source: §6 · p. 8

7 A Research Agenda

The agenda proposes stronger AMM evaluation, testable principles of memory utility, and optimization methods that combine parametric and non-parametric updates. It also calls for studying agent information needs while keeping humans in control.

7.1 Toward Systematic Evaluation Protocols

  • Benchmarks should test constants, functions, and operators separately and in composition, alongside learning and forgetting from misleading or contradictory episodes and effective deletion.Most current benchmarks focus on using preconstructed stores; learning and forgetting need first-class tasks.Source: §7.1 · p. 8
  • Benchmarks should include task outcomes, corrections, and delayed rewards; test cross-user and cross-session leakage; and report results across two backbones with token, latency, and storage costs.These desiderata make optimization, scope, and cost–quality trade-offs measurable.Source: §7.1 · pp. 8–9

7.2 Reconnecting to Established IR

  • Table 5 links open AMM questions to established IR research, but its mappings are starting points because established solutions often do not transfer unchanged.Related IR work on ranking losses may inform AMM research on deduplication, temporal validity, and feedback scope.Source: §7.2 · p. 9, Table 5 · p. 9

7.3 Toward a Theory of Memory Utility

  • AMM lacks both a practical end-to-end learning route and principled conditions for memory designs, motivating hypotheses with observable consequences and tests.Systems combine prompts, stores, policies, and foundation models that most researchers cannot retrain; current designs often rely on human-memory analogies without specifying when they help.Source: §7.3 · p. 9
  • AMM needs testable hypotheses: prioritize memories by expected future reward net of costs, balance abstraction’s transfer benefits against misapplication risk, and apply corrections at the narrowest supported scope.These are conjectures; expected utility depends on an unknown future task distribution, abstraction may suit repeated rather than heterogeneous tasks, and scope should widen with repeated independent evidence.Source: §7.3 · pp. 9–10

7.4 Hybrid Parametric and Non-Parametric Optimization

  • Research should characterize convergence of textual revision, improve credit assignment for delayed outcomes, develop principled joint parametric and non-parametric optimization, and optimize memory structure as well as content.The agenda also asks when knowledge belongs in weights versus text, and how consolidation can be reversed.Source: §7.4 · p. 10

7.5 Studying the Agent, and Keeping the Human in the Loop

  • Studying agent interaction logs and memory presentation can reveal information needs, while human control requires meaningful ability to inspect, correct, and delete memories, including after consolidation.The agenda also identifies memory poisoning and membership inference as emerging concerns that require more than systems work.Source: §7.5 · p. 10

8 Conclusion

The paper concludes that agentic memory management is a new form of IR problem and that IR can help build its foundations. It calls for systematic evaluation, testable utility principles, hybrid optimization, and human involvement.

8 Conclusion

  • AMM extends established IR activities by adding the agent’s use of information, with memory management supporting assistants that improve with use rather than starting afresh each session.The activities of structuring, analyzing, organizing, storing, searching, and retrieving remain relevant as the information consumer changes.Source: §8 · p. 10
  • AMM is a new form of IR problem, and IR can help build stronger foundations through systematic evaluation, testable utility principles, hybrid optimization, and human involvement.The paper frames the next memory to organize as the one inside the agent.Source: §8 · p. 10
Loading 2610.32464v1…