Computation and Language · arXiv:2602.16313 v2 · submitted 18 Feb 2026
MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks
The paper says
“MemoryArena supports evaluation across web navigation, preference-constrained planning, progressive information search, and sequential formal reasoning, and reveals that agents with near-saturated performance on existing long-context memory benchmarks like LoCoMo perform poorly in our agentic setting, exposing a gap in current evaluations for agents with memory.”
from the abstract · picked out as the main claim by a classifier
Reviews
Reviews by people
No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Existing evaluations of agents with memory typically assess memorization and action in isolation. One class of benchmarks evaluates memorization by testing recall of past conversations or text but fails to capture how memory is used to guide future decisions. Another class focuses on agents acting in single-session tasks without the need for long-term memory. However, in realistic settings, memorization and action are tightly coupled: agents acquire memory while interacting with the environment, and subsequently rely on that memory to solve future tasks. To capture this setting, we introduce MemoryArena, a unified evaluation gym for benchmarking agent memory in multi-session Memory-Agent-Environment loops. The benchmark consists of human-crafted agentic tasks with explicitly interdependent subtasks, where agents must learn from earlier actions and feedback by distilling experiences into memory, and subsequently use that memory to guide later actions to solve the overall task. MemoryArena supports evaluation across web navigation, preference-constrained planning, progressive information search, and sequential formal reasoning, and reveals that agents with near-saturated performance on existing long-context memory benchmarks like LoCoMo perform poorly in our agentic setting, exposing a gap in current evaluations for agents with memory. MemoryArena is now released at https://memoryarena.github.io/.
The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23 · no language-model review
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.15 · states limitation 0.14 · human participants 0.29 · controlled comparison 0.34 · randomized assignment 0.01 · states quantitative result 0.06 · compares with prior results 0.93 · mentions code or data availability 0.95
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v2). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Existing evaluations of agents with memory typically assess memorization and action in isolation. | background 0.78 | 0.04 | 0.03 | asserted 1.00 |
| One class of benchmarks evaluates memorization by testing recall of past conversations or text but fails to capture how memory is used to guide future decisions. | background 0.53 | 0.04 | 0.03 | asserted 1.00 |
| Another class focuses on agents acting in single-session tasks without the need for long-term memory. | background 0.81 | 0.03 | 0.02 | asserted 1.00 |
| However, in realistic settings, memorization and action are tightly coupled: agents acquire memory while interacting with the environment, and subsequently rely on that memory to solve future tasks. | background 0.56 | 0.07 | 0.07 | asserted 1.00 |
| To capture this setting, we introduce MemoryArena, a unified evaluation gym for benchmarking agent memory in multi-session Memory-Agent-Environment loops. | contribution 0.99 | 0.92 | 0.08 | asserted 1.00 |
| The benchmark consists of human-crafted agentic tasks with explicitly interdependent subtasks, where agents must learn from earlier actions and feedback by distilling experiences into memory, and subsequently use that memory to guide later actions to solve the overall task. | method 0.80 | 0.80 | 0.11 | asserted 1.00 |
| MemoryArena supports evaluation across web navigation, preference-constrained planning, progressive information search, and sequential formal reasoning, and reveals that agents with near-saturated performance on existing long-context memory benchmarks like LoCoMo perform poorly in our agentic setting, exposing a gap in current evaluations for agents with memory. | result 0.94 | 0.94 | 0.79 | asserted 1.00 |
| MemoryArena is now released at https://memoryarena.github.io/. | contribution 0.91 | 0.44 | 0.02 | asserted 0.99 |
What kind of paper the classifier reports this is: dataset or benchmark 1.00 · evidence from observation or experiment 0.94
What was examined, and what was not
Sources read
- arXiv listing abstract v2
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,302 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Cited by 21 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- MOSAIC: An Omni-Modal Benchmark for Long-Term Memory and Cross-Modal Reasoning in Conversational Agents cs.CL · Oct 2026
- Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents cs.AI · Oct 2026
- The Right Fact, the Wrong Decision: SciValidityBench for Scientific Agent Memory cs.AI · Oct 2026
- Continual Learning Exam: Benchmarking Long-Horizon RSI and Experiential Learning in LLM Agents cs.LG · Oct 2026
- HABIT-Bench: Evaluating Habit Induction from Longitudinal Weak Evidence in Agent Memory cs.AI · Oct 2026
- Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms cs.AI · Sep 2026
- MemMend: A Selective Revision Framework for Memory Question Answering cs.AI · Nov 2026
- Where Should LLM Agents Adapt at Test Time? In Context, in Weights, or Both cs.LG · Nov 2026
- Towards Robust Long-Lived Agents: Benchmarking Experience Lifecycles in Evolving Environments cs.AI · Nov 2026
- MemSteer: Turning User Histories into Supervision for Agent Memory Optimization cs.AI · Oct 2026
- 4DMemory: Benchmarking Long-Horizon Spatiotemporal Memory with Harness-Augmented Agents cs.CV · Oct 2026
- Memorizing Who, Knowing How: Benchmarking Personal and Procedural Memory for Persistent Assistants cs.AI · Oct 2026
- When Agent Memory Needs to Compute: Evidence Access and Executable Aggregation cs.AI · Oct 2026
- MAGMA-Bench: Evaluating Long-Horizon Embodied Agents under Dynamic Tasks and Execution Failures cs.AI · Oct 2026
- Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies cs.AI · Oct 2026
- Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments? cs.AI · Oct 2026
- When Should Agents Check External State? Budgeting Observations for Stored Intentions cs.AI · Sep 2026
- FlowState: Execution State as Memory for Long-Horizon LLM Agents cs.AI · Sep 2026
- KV-LTM: Persistent Per-Session KV Caches as Long-Term Memory for Personal Assistants cs.AI · Nov 2026
- Who Schedules Agent Memory? A Sidecar Approach to Long-Horizon Tasks cs.AI · Oct 2026
- CTXAgent - Self-Controlled Context Management of LLM Agents cs.AI · Oct 2026
How the literature uses it
51 arXiv papers cite this in 90 sentences
Papers citing it · 55 sentences without a confident label
Contests it 1 paper
-
Early long-term conversational-memory suites are text-only and omit visual modalities altogether [ref] ; subsequent multimodal benchmarks add images but treat them as static snapshots and ignore document-level artifacts [ref] , while document-QA benchmarks probe rich multimodal reasoning but restrict it to a single document within a single turn [ref] .
M$^3$Exam: Benchmarking Multimodal Memory for Realistic User-Agent Interactions · Introduction · contest 0.71
Builds on it 1 paper
-
We instantiate it as CMI-Travel, a controlled extension of the Travel Planner tasks in MemoryArena [ref] .
Controlled Memory Interference in Continual LLM Agents · Introduction · 0.99
Uses its method or tool 3 papers
-
The MemoryArena environments and evaluators are available from the benchmark authors [ref] .
MemoryLake on MemoryArena: A Matched Study of Agent Memory Backends · Code and Data Availability · 0.97
-
We then follow the simulated user-in-the-loop protocol used in PersonaGym [ref] to evaluate on ClawArena [ref] and MemoryArena [ref] .
TellOnce: Correcting Coding Agents Once, Improving Their Future Actions · Introduction · 0.81
-
We additionally use MemoryArena to study inter-session dependencies [ref] .
Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads · Workload Suite · 0.92
Uses its data or values 2 papers
-
Vector RAG represents a simple embedding-based retrieval pipeline; among methods in the original MemoryArena table, text-embedding-3-small had the highest reported suite-level average SR [ref] .
MemoryLake on MemoryArena: A Matched Study of Agent Memory Backends · Systems · 0.97
-
We apply CMI to state-update cases derived from TravelPlanner in MemoryArena [ref] , PERMA [ref] , LongMemEval [ref] , HorizonBench [ref] , and the Tracking task from MEME [ref] .
Controlled Memory Interference in Continual LLM Agents · Datasets and Protocols · 0.87
Cites it as support 3 papers
-
Systems that score near-perfectly on conversational recall benchmarks can drop to 40–60% when the same facts must be used inside a downstream task such as web navigation or constrained planning [ref] ; passive recall and active use are not the same skill. * Memory must update.
Beyond Memory Leaderboards: Evaluating Scientific Memory as Budgeted Context Restoration · Introduction · 0.96
-
MemoryArena [ref] and large environment suites like AI Gamestore [ref] and ARC-AGI-3 [ref] point toward persistent state and cross-session dependencies [ref] .
Interactive Evaluation Requires a Design Science · Axis 1: Evaluation Inputs · 0.98
-
MemoryArena [ref] shows that models near-perfect on recall benchmarks drop sharply when memory must drive actions, motivating evaluation beyond passive recall.
Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents · Related Work · 0.97
Cites it as background 21 papers
-
Agent-memory and personalization benchmarks instead give the model a pre-written long-term memory and measure how it is recalled, reused, or over-applied when answering, including cross-domain leakage, over-personalization, and memory-induced sycophancy [ref] .
Agents Don't Just Agree, They Remember: Benchmarking Persistent Sycophancy in Self-Improving Personal Agents · Introduction · 0.87
-
Recent benchmarks have made important progress by defining realistic long-horizon memory tasks, including LongMemEval [ref] , MemoryAgentBench [ref] , MemoryArena [ref] , and BEAM [ref] .
MemDelta: Controlled Baselines and Hidden Confounds in Agent Memory Evaluation · Introduction · 0.93
-
Some probe long-horizon recall and multi-session reasoning, including updating stale facts and forgetting what is outdated [ref] , while others probe how closely an assistant's responses match a user's profile, preferences, and history [ref] .
MemAudit: Auditing Long-Term Agent Memory via Hidden User-State Recovery · Introduction · 0.83
What it cites · 10 sentences without a confident label
Uses the method or tool of 3 papers
-
We include four mainstream agents with external memory: MemGPT [ref] , Mem0 and its graph version Mem0-g [ref] ,
Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory · Experimentation Setup · 0.96
-
We consider different retrieval methods, including BM25, an embedding-based RAG method that retrieves based on semantic similarity (using OpenAI ), and two structured RAG approaches, MemoRAG [ref] and GraphRAG [ref] , in our evaluation.
From Local to Global: A Graph RAG Approach to Query-Focused Summarization · Experimentation Setup · 0.99
-
and ReasoningBank [ref] .
ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory · Experimentation Setup · 0.91
Uses data or values from 1 paper
-
Our test data builds upon BrowseComp-Plus [ref] .
BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent · Task Composition and Data Preparation · 0.91
Cites as background 11 papers
-
LoCoMo [ref] , LongMemEval [ref] , MemoryAgentBench [ref] , MemoryBench [ref] , and EvoMem [ref] assess whether models can retain and recall information introduced in the previous interactions.
Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory · Related Works · 0.94
-
LoCoMo [ref] , LongMemEval [ref] , MemoryAgentBench [ref] , MemoryBench [ref] , and EvoMem [ref] assess whether models can retain and recall information introduced in the previous interactions.
MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems · Related Works · 0.94
-
through recall or retrieval over static long-context inputs in question answering or summarization settings [ref] , including benchmarks such as LoCoMo [ref] and LongMemEval [ref] .
Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions · Introduction · 0.95
Challenges and discussion
Know something about this paper? Leave a comment.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment