AuthorsJing Yu, Yibo Zhao, Jiaming Zhang, Xiang Li
TL;DR
LazyMem defers all lossy memory construction to query time with overlapping window compression, reaching 0.85 LJ on LongMemEval using only 213 memory tokens.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Dialogue histories quickly exceed the context window, and useful evidence is sparse and scattered across verbose conversations, making retrieval difficult.
When agents broaden recall to improve coverage, they flood downstream reasoning with noise, while write-time compression irreversibly discards details future queries may need.
HOW IT WORKS
LazyMem’s core mechanism chains Hybrid Retrieval, History Windowing, and a lightweight Memory-Processing Model to build query-conditioned compact memory.
You can think of LazyMem like disk plus RAM: all raw conversations stay on disk, and only query-relevant slices are loaded and compressed into a small RAM-like context.
This design lets LazyMem retrieve broadly for coverage but surface only compressed, useful evidence, avoiding permanent write-time loss that a plain context window cannot prevent.
DIAGRAM
This diagram shows how LazyMem turns retrieved raw messages into compact, query-conditioned memory via overlapping windows and KEEP DROP decisions.
DIAGRAM
This diagram shows how LazyMem trains its memory-processing model using teacher supervision and group-based reinforcement learning with action and quality rewards.
PROCESS
Retrieve broadly
LazyMem runs Hybrid Retrieval with dense and BM25 retrievers plus Reciprocal Rank Fusion and a cross-encoder reranker to build a high-recall candidate pool Rq of up to 50 messages.
History windowing
LazyMem applies History Windowing, expanding each retrieved hit with w neighboring messages, merging overlaps, and splitting long spans into overlapping sub-windows capped at L messages.
Query-conditioned memory construction
For each window, the Memory-Processing Model predicts KEEP DROP per message and rewrites kept messages into compressed, query-relevant content, then deduplicates across overlapping windows.
Answer generation
LazyMem concatenates compressions into compact memory mq and feeds mq plus the query to the Answer Model, which produces the final prediction under a fixed reasoning prompt.
KEY CONTRIBUTIONS
Lazy retrieve-then-construct memory system
LazyMem preserves raw interactions, retrieves broadly at query time, and constructs compact evidence via overlapping windows, avoiding irreversible write-time loss while bounding context length and latency.
4B memory-processing model with SFT plus GRPO
LazyMem trains a 4B Memory-Processing Model using supervised fine-tuning and format-gated GRPO with action and quality rewards, boosting LongMemEval LJ from 0.41 to 0.85.
Efficient long-term memory with strong generalization
LazyMem-4B reaches 0.85 LJ on LongMemEval using 213 memory tokens and 0.68 LJ on LoCoMo without target-domain training, reducing answer-context tokens by 68.7× versus RAG Top50.
RESULTS
LongMemEval LJ
0.85 score
+0.03 over StructMem
LoCoMo LJ
0.68 score
+0.11 over NanoMemory
Answer-context memory tokens
213 tokens
21.0× fewer than StructMem
Latency on LongMemEval
40.86 s
14.49 s faster than NanoMemory
On LongMemEval, which tests multi-session and temporal reasoning over long histories, LazyMem-4B attains 0.85 LLM-judge accuracy with only 213 answer-context memory tokens. On LoCoMo, which stresses multi-hop and temporal memory, LazyMem-4B reaches 0.68 LJ without target-domain training, showing that query-time construction scales beyond its training corpus.
BENCHMARK
On LongMemEval, which tests multi-session and temporal reasoning over long histories, LazyMem-4B attains 0.85 LLM-judge accuracy with only 213 answer-context memory tokens. On LoCoMo, which stresses multi-hop and temporal memory, LazyMem-4B reaches 0.68 LJ without target-domain training, showing that query-time construction scales beyond its training corpus.
BENCHMARK
Overall LLM-judge accuracy (LJ) on the LongMemEval test split.
BENCHMARK
Overall LLM-judge accuracy (LJ) on LoCoMo categories 1–4 from the last two conversations.
KEY INSIGHT
LazyMem-4B, a 4B memory-processing model, matches or beats several systems that use Qwen3-32B for memory, reaching 0.85 LJ on LongMemEval.
This is surprising because builders often assume larger memory models are necessary for high-quality selection and compression, yet LazyMem shows training and architecture can offset scale.
WHY IT MATTERS
LazyMem unlocks long-term agents that can safely store raw histories and still answer queries using small, high-utility memory slices at inference time.
Builders can now design agents that retrieve broadly, construct selectively, and keep latency and context budgets manageable without sacrificing multi-session reasoning quality.
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…