AuthorsChengzhi Liu, Yuzhe Yang, Sophia Xiao Pu et al.
TL;DR
WorldMemArena reframes agent memory as an Action World Interaction Loop with a four-stage lifecycle and exposes that high memory recall (e.g., 86.83%) often fails to translate into strong QA correctness (e.g., 50.14%).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory benchmarks collapse performance into single QA accuracy, ignoring whether agents correctly write, update, retrieve, and use evidence over time.
On long-horizon multimodal tasks, agents with high storage quality still fail to act reliably, because WorldMemArena shows that better memory writing and storage do not guarantee better performance.
HOW IT WORKS
WorldMemArena defines multimodal agent memory through an Action World Interaction Loop with four stages: Observe to Write, Update and Consolidate, Retrieve for Decision, and Use and Act operating over a Memory Library.
Think of Observe to Write and Update and Consolidate as RAM that selectively logs and revises state, while Retrieve for Decision and Use and Act behave like a card catalog plus a decision engine that surfaces and applies the right entries.
This lifecycle lets WorldMemArena test capabilities that a plain context window cannot, by separately scoring writing quality, update handling, retrieval coverage, and downstream QA correctness across Lifelong Evolution and Agentic Execution regimes.
DIAGRAM
This diagram shows how WorldMemArena models each session as an Action World Interaction Loop over latent world state, observations, actions, feedback, and memory.
DIAGRAM
This diagram shows how WorldMemArena constructs multi-session tasks, annotates gold memory points, and evaluates the four lifecycle stages.
PROCESS
Observe to Write
WorldMemArena checks whether agents, given a session trajectory, can produce a selective memory delta that matches gold memory points instead of logging the full trajectory.
Update and Consolidate
WorldMemArena evaluates how newly written memory is integrated, requiring correct handling of state updates and removal or overwrite of obsolete entries for temporal consistency.
Retrieve for Decision
WorldMemArena measures whether agents can retrieve decision-relevant evidence for a query, scoring coverage and NDCG against annotated evidence chains from the Memory Library.
Use and Act
WorldMemArena tests whether retrieved evidence is faithfully used to answer checkpoint questions, across Basic, Robustness, Reasoning, and Multimodal axes with LLM-as-a-Judge, F1, and BLEU.
KEY CONTRIBUTIONS
Action World Interaction Loop formulation
WorldMemArena formulates multimodal agent memory as an Action World Interaction Loop with four observable stages, enabling stage-level diagnosis instead of single end-of-task accuracy.
WorldMemArena multi-session multimodal benchmark
WorldMemArena provides 461 multi-session tasks, 24,258 QA pairs, and 15,595 images across Lifelong Evolution and Agentic Execution, annotated with gold memory points, updates, distractors, and evidence chains.
Unified comparison of memory paradigms
WorldMemArena conducts head-to-head evaluation of long-context agents, manually designed external memory and RAG systems, and harness-based memory agents, revealing instability and multimodal bottlenecks.
RESULTS
Recall
86.83%
+2.27 over Qwen3-VL-Embedding-8B (84.56% Recall)
QA-C
54.63%
+15.01 over UniversalRAG (39.62% QA-C)
Update
58.86%
+0.88 over Qwen3-VL-Embedding-8B (57.98% Update)
RC
74.19%
+13.26 over UniversalRAG (60.93% RC retrieval coverage)
On WorldMemArena’s lifecycle evaluation, A-Mem and M2A illustrate that high memory Recall (up to 86.83%) and strong retrieval coverage (74.19% RC) do not automatically yield perfect QA correctness. These numbers prove that WorldMemArena can separate storage quality from effective evidence use, exposing gaps between memory writing, update handling, retrieval, and final QA performance.
BENCHMARK
On WorldMemArena’s lifecycle evaluation, A-Mem and M2A illustrate that high memory Recall (up to 86.83%) and strong retrieval coverage (74.19% RC) do not automatically yield perfect QA correctness. These numbers prove that WorldMemArena can separate storage quality from effective evidence use, exposing gaps between memory writing, update handling, retrieval, and final QA performance.
BENCHMARK
QA-C (QA Correct) comparison across representative external memory and RAG systems on WorldMemArena.
KEY INSIGHT
WorldMemArena shows that Qwen3-VL-Embedding-8B reaches 86.22% memory Recall but only 51.86% QA-C, while UniversalRAG has 84.56% Recall yet just 39.62% QA-C.
This breaks the assumption that more and better stored memories automatically improve answers, revealing that retrieval and use stages are the true bottlenecks.
WHY IT MATTERS
WorldMemArena unlocks the ability to pinpoint whether an agent’s failure comes from writing, maintenance, retrieval, or use, instead of blaming generic memory or context length.
Builders can now design and debug memory systems that revise stale state, resist distractors, and truly reuse multimodal experience across sessions, rather than just extending context windows.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…