AuthorsRuizhe Li, Mingxuan Du, Benfeng Xu, Zhendong Mao
TL;DR
Keep It InMind shows that query-conditioned retrieval creates an implicit-association blind spot, with 84.0% in-context indirect accuracy vs at most 16.0% under retrieval.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Keep It InMind shows that agents answer 84.0% of indirect queries when the decisive memory is in context, but only up to 16.0% when it must be retrieved.
This implicit-association blind spot means systems like HippoRAG 2, A-Mem, xMemory, and A-RAG recall facts on demand yet fail to apply them, recommending unsafe answers such as macarons to users with tree nut allergy.
HOW IT WORKS
Keep It InMind builds the InMind benchmark with 125 tasks and paired controls: Naive query, Indirect query, and Target recall to separate storage, knowledge, and retrieval failures.
Think of InMind as a stress test where a long-term memory store is disk, the always-visible profile is RAM, and the indirect query is a safety-critical interrupt that must see the right facts.
This design lets Keep It InMind pinpoint that the query-conditioned retrieval interface, not storage or backbone knowledge, causes the 84.0% to 16.0% collapse on implicit associations.
DIAGRAM
This diagram shows how Keep It InMind tests systems with a naive direct query and then an indirect query to separate storage from application.
DIAGRAM
This diagram shows how Keep It InMind injects memories into LME-s, then evaluates six memory systems plus controls on naive, recall, and application metrics.
PROCESS
Task Extraction
Keep It InMind uses task extraction to build a user fact, naive query, indirect query, and source-grounded bridge from public chunks, enforced by strict prompts.
Similarity Filtering
Keep It InMind applies BM25 and MiniLM similarity filtering so the decisive memory has no lexical or dense cue to the indirect query, ensuring true implicit associations.
Conflict and Expert Verification
Keep It InMind runs conflict checks against the LME-s persona and expert verification to confirm each bridge is factual, answer-changing, and free of overt retrieval cues.
Evaluation Protocol
Keep It InMind injects the memory into LME-s, lets 38 sessions pass, then issues naive and indirect queries to systems, scoring Naive recall, Target recall, and Application.
KEY CONTRIBUTIONS
Defining the implicit-association blind spot
Keep It InMind formalizes the retrieval hypothesis and shows it fails on implicit associations where relevance is created by a knowledge bridge absent from both memory and query.
InMind benchmark with paired controls
Keep It InMind introduces InMind, a 125-task benchmark with 113 source-grounded tasks and paired Naive, In-context backbone, and Target recall controls to separate storage, knowledge, and retrieval failures.
Locating failure in query-conditioned interfaces
Keep It InMind shows six memory systems recall facts on demand at up to 100.0% yet reach at most 14.4% indirect application, while a minimal always-in-state probe recovers 68.8%.
RESULTS
Naive recall
100.0%
+16.0 over MemoryOS emb3-large
Target Recall
12.0%
A-Mem with text-embedding-3-large
Application
14.4%
MemoryOS with text-embedding-3-large on InMind
Backbone indirect
84.0%
vs best retrieval configuration at 16.0% on InMind
Keep It InMind evaluates six memory systems plus Naive RAG and a backbone control on the InMind benchmark. The 84.0% backbone indirect accuracy vs at most 16.0% retrieval application proves that query-conditioned interfaces, not storage or backbone knowledge, cause most failures on implicit associations.
BENCHMARK
Keep It InMind evaluates six memory systems plus Naive RAG and a backbone control on the InMind benchmark. The 84.0% backbone indirect accuracy vs at most 16.0% retrieval application proves that query-conditioned interfaces, not storage or backbone knowledge, cause most failures on implicit associations.
BENCHMARK
Indirect application (%) for backbone vs representative retrieval systems on the 125-task InMind benchmark.
KEY INSIGHT
Keep It InMind finds that six memory systems recall injected facts on demand at up to 100.0% naive accuracy yet apply them under indirect queries at no more than 14.4%.
This is counterintuitive because users and builders assume that correct direct recall implies correct application, but Keep It InMind shows a 68.0 percentage-point gap between the 84.0% backbone and the 16.0% best retrieval configuration.
WHY IT MATTERS
Keep It InMind gives the field a precise diagnostic for when long-term memory fails silently, separating storage, backbone knowledge, and retrieval interface errors on implicit associations.
With InMind and the always-in-state probe, builders can now measure and target routing decisions—what must stay visible in state versus live only in retrieval—rather than just improving similarity search or storage capacity.
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…