AuthorsChaoqun Zhan, Qiang Zhou, Guannan Li et al.
TL;DR
MemoryLake uses structured multi-track memory with prioritized conclusions and evidence retrieval to reach a 20.5% macro-average SR vs 13.6% best baseline on MemoryArena.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Most agent memory benchmarks focus on post hoc recall, not whether remembered information actually changes later actions or supports multi step objectives.
MemoryArena exposes that agents can recall facts yet still fail end to end tasks, leaving multi session, interdependent workloads under evaluated and poorly understood.
HOW IT WORKS
MemoryLake’s core mechanism separates confirmed conclusions, supporting evidence, and reusable experience into parallel tracks with different presence policies, all orchestrated through a shared gpt-5-mini backbone and bge-m3 dense retrieval.
You can think of MemoryLake like a computer with fast RAM for conclusions, a searchable disk for evidence, and a library of reusable procedures acting as cached skills.
This design lets MemoryLake keep critical conclusions always present, pull detailed evidence on demand, and reuse prior reasoning in ways a flat context window or naive vector RAG cannot.
DIAGRAM
This diagram shows how MemoryLake handles writes after each subtask and assembles memory before the next subtask in MemoryArena.
DIAGRAM
This diagram shows the matched evaluation protocol comparing MemoryLake, Mem0, Vector RAG, and Long Context across five MemoryArena domains.
PROCESS
Matched system level protocol
MemoryLake joins the shared MemoryArena agent framework with the gpt-5-mini alias and fixed task IDs, ensuring only the memory backend differs from Mem0, Vector RAG, and Long Context.
Memory write after each subtask
After every completed subtask, MemoryLake writes the full trajectory and outcome into its multi track store, populating confirmed conclusions, supporting evidence, and reusable experience.
Presence policies for heterogeneous content
MemoryLake prioritizes confirmed conclusions for future prompts, stores supporting evidence for bge-m3 dense retrieval, and consolidates reusable experience for transfer across subtasks.
Bounded memory assembly and fallback
Before each new subtask, MemoryLake assembles a bounded context from its tracks, and if the internal memory service fails, it returns an empty context so the agent can continue.
KEY CONTRIBUTIONS
Matched system level comparison
MemoryLake is evaluated as a complete backend configuration against Mem0, Vector RAG, and Long Context across all five MemoryArena domains, using identical gpt-5-mini agents and task IDs.
End to end SR and process metrics
MemoryLake reports both per domain SR and process metrics like PS, SPS, and step match, plus a 20.5% macro average SR compared to a 13.6% best comparator.
Workload level interpretation and transparency
MemoryLake provides a workload dependent analysis of memory backends and documents sampling, judge dependence, and Mem0’s capped re run to clarify reproducibility limits.
RESULTS
Math PS SR
29.5% PS / 22.5% SR
+10.0pp SR over Long Context
Physics PS SR
64.5% PS / 60.0% SR
+15.0pp SR over Mem0
Progressive PS SR
6.7% PS / 20.0% SR
+10.0pp SR over Vector RAG
Macro Avg SR
20.5%
+6.9pp over 13.6% Long Context
These metrics come from the shared MemoryArena evaluation sets, covering formal mathematics, physics, group travel planning, bundled web shopping, and progressive retrieval. The 20.5% macro average SR shows that MemoryLake achieves higher end to end success than the 13.6% best comparator under a matched agent and model configuration.
BENCHMARK
These metrics come from the shared MemoryArena evaluation sets, covering formal mathematics, physics, group travel planning, bundled web shopping, and progressive retrieval. The 20.5% macro average SR shows that MemoryLake achieves higher end to end success than the 13.6% best comparator under a matched agent and model configuration.
BENCHMARK
Equal weight macro average SR across five MemoryArena domains.
KEY INSIGHT
In bundled web shopping, all four systems achieve only one bundle level success out of 150, with the sole success belonging to the Long Context baseline.
This is surprising because MemoryLake’s structured memory and Vector RAG’s retrieval achieve similar step match rates around 30%, yet none translate that into higher end to end SR under strict bundle success criteria.
WHY IT MATTERS
MemoryLake shows that structured multi track memory can shift end to end success on interdependent tasks, especially in formal reasoning and progressive retrieval.
Builders can now treat memory backends as workload specific design choices, using MemoryLake style presence policies when exact conclusion reuse and aggregation matter, rather than relying solely on longer context windows or naive vector RAG.
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…