AuthorsDongfang Li, Zixuan Liu, Junmai Wang et al.
TL;DR
LycheeMemory V2 uses semantic segment-level consolidation with typed records to reach 89.22% on LoCoMo and 92.20% on LongMemEval-S while cutting construction tokens by up to 86.0% versus A-Mem.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory systems rely on eager consolidation, invoking LLMs after each interaction, so construction tokens reach 1459.9K for A-Mem on LoCoMo.
This makes long-horizon LLM agents like A-Mem and Mem0 increasingly costly as conversations grow, forcing a trade-off between accuracy, token budget, and usable long-term memory.
HOW IT WORKS
LycheeMemory V2 centers on Online Semantic Segmentation, Segment-Level Memory Encoding, Structured Evidence Organization, and Plan-Guided Multi-Route Retrieval to build memory from coherent dialogue segments.
Think of LycheeMemory V2 as turning a raw chat log into a library card catalog, where each card is a typed record linked by entities, topics, and time.
This design lets LycheeMemory V2 preserve fine-grained evidence and temporal relations at segment granularity, instead of overstuffing a single context window or paying for turn-level consolidation.
DIAGRAM
This diagram shows how LycheeMemory V2 plans and executes multi-route retrieval from its structured memory store when answering a user query.
DIAGRAM
This diagram shows how LycheeMemory V2 is evaluated on LoCoMo and LongMemEval-S, including construction-side and retrieval-side ablations.
PROCESS
Online Semantic Segmentation
LycheeMemory V2 computes semantic surprise and cohesion scores to decide when buffered exchanges form a coherent segment worth consolidating.
Segment-Level Memory Encoding
LycheeMemory V2 encodes each finalized segment into typed, self-contained records with entities, topics, temporal scopes, and provenance, plus disambiguation feedback.
Structured Evidence Organization
LycheeMemory V2 inserts records into a vector store and builds entity, topic, entity topic, temporal, and event frame indexes for structured access.
Plan-Guided Multi-Route Retrieval
LycheeMemory V2 uses a query planner to define recall routes, then executes direct record, structured node, temporal, and raw turn recall before fusion and answer generation.
KEY CONTRIBUTIONS
Semantic Segment-Level Memory Construction
LycheeMemory V2 introduces Online Semantic Segmentation plus Segment-Level Memory Encoding, reducing construction tokens on LoCoMo from 1459.9K for A-Mem to 204.1K while improving accuracy.
Typed Records with Structured Evidence Organization
LycheeMemory V2 encodes segments into typed records and organizes them via entity, topic, temporal, event frame, and entity topic indexes for richer retrieval than flat RAG.
Plan-Guided Multi-Route Retrieval Under Token Budgets
LycheeMemory V2 uses Plan-Guided Multi-Route Retrieval to reach 92.20% accuracy on LongMemEval-S while keeping query tokens at 8.88K, lower than A-Mem's 15.46K.
RESULTS
LoCoMo Overall Accuracy
89.22%
+20.39 percentage points over A-Mem (68.83%) with GPT-4.1-Mini
LongMemEval-S Overall Accuracy
92.20%
+20.60 percentage points over A-Mem (71.60%) with GPT-4.1-Mini
LoCoMo Construction Tokens
204.1K
-86.0% versus A-Mem (1459.9K) construction tokens using GPT-4.1-Mini
LongMemEval-S Construction Tokens
304.7K
-75.9% versus A-Mem (1264.3K) construction tokens using GPT-4.1-Mini
LoCoMo tests single hop, multi hop, temporal, and open domain long-term memory questions, while LongMemEval-S stresses user facts, preferences, multi session reasoning, knowledge updates, and temporal reasoning.
The main results show LycheeMemory V2 can substantially boost long-term QA accuracy on both benchmarks while sharply reducing construction tokens compared to A-Mem and TiMem.
BENCHMARK
LoCoMo tests single hop, multi hop, temporal, and open domain long-term memory questions, while LongMemEval-S stresses user facts, preferences, multi session reasoning, knowledge updates, and temporal reasoning. The main results show LycheeMemory V2 can substantially boost long-term QA accuracy on both benchmarks while sharply reducing construction tokens compared to A-Mem and TiMem.
BENCHMARK
Overall accuracy (%) on LoCoMo for LycheeMemory V2 and representative memory baselines.
BENCHMARK
Overall accuracy (%) on LongMemEval-S for LycheeMemory V2 and representative memory baselines.
KEY INSIGHT
LycheeMemory V2 cuts construction tokens on LoCoMo by 86.0% compared to A-Mem, yet increases overall accuracy from 68.83% to 89.22%.
This is surprising because cheaper memory is often assumed to be coarser and less accurate, but LycheeMemory V2 shows granularity and consolidation strategy matter more than sheer token volume.
WHY IT MATTERS
LycheeMemory V2 makes it practical to run long-horizon LLM agents with rich, queryable memories under tight construction and query token budgets.
Builders can now deploy agents that remember multi session events, preferences, and temporal updates across 100K token histories without resorting to full context replay or expensive turn-level consolidation.
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…