AuthorsZhengjun Huang, Zhoujin Tian, Qintian Guo et al.
TL;DR
LiCoMemory uses the hierarchical CogniGraph with hierarchy temporal rerank to reach 73.80% accuracy on LongMemEval, +9.0 points over Mem0g.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing graph based memory like GraphRAG can require up to 20 minutes for graph construction and over 2 minutes of query latency per dialogue.
These heavy, flat graphs mix semantics and topology, causing redundant nodes, scattered retrieval, and degraded long term reasoning in conversational agents.
HOW IT WORKS
LiCoMemory centers on CogniGraph, Query Processing and Integrated Rerank, and Real Time Interactions to decouple storage from semantic indexing across three layers.
You can think of CogniGraph as a card catalog for memory, where summaries are shelves, triples are index cards, and chunks are the books on the back wall.
This design lets LiCoMemory navigate from high level summaries to timestamped triples and raw chunks, enabling temporally aware reasoning far beyond a plain context window.
DIAGRAM
This diagram shows how LiCoMemory processes a user query, retrieves through CogniGraph, and then performs real time memory updates.
DIAGRAM
This diagram shows how LiCoMemory is evaluated on LongMemEval and LoCoMo, including ablations of structured retrieval, temporal awareness, and summaries.
PROCESS
CogniGraph: A Lightweight and Semantically Aware Graph Structure
LiCoMemory builds CogniGraph with session level summaries, entity relation triples, and chunk level dialogue, linking them via unique identifiers for hierarchical indexing.
Query Processing and Integrated Rerank
LiCoMemory extracts entities, matches them to session summaries, retrieves triples, and computes Ssem and R(t) using the harmonic mean and Weibull based temporal decay.
Real Time Interactions
LiCoMemory retrieves evidence for the agent response, then incrementally updates session summaries, performs triple extraction and deduplication, and links new chunks.
Hierarchy and Temporally Sensitive Retrieval
LiCoMemory performs top down retrieval from summaries to triples to chunks, guided by hierarchical structure and temporal relevance to maintain coherent answers.
KEY CONTRIBUTIONS
CogniGraph for semantic organization
LiCoMemory introduces CogniGraph with session level summaries, entity relation triples, and chunk level dialogue, decoupling storage from indexing and enabling lightweight updates.
Hierarchy and temporally sensitive retrieval
LiCoMemory combines session level relevance Ss, triple level relevance St, and a Weibull based decay w(Δτ) to compute R(t) for temporally aware retrieval.
Efficient and real time memory operations
LiCoMemory reduces construction latency on LoCoMo from 1772s in Mem0 to 21s and cuts construction tokens from 49.3k to 13.52k while improving accuracy.
RESULTS
Acc.
73.80%
+9.0 over Mem0g
Rec.
76.63%
+7.10 over Mem0g
TR
1.61s
0.14s faster than Mem0g on LongMemEval with GPT 4o mini
KR
1.7k
1.2k fewer tokens than Zep on LongMemEval with GPT 4o mini
On LongMemEval, which tests single session, multi session, temporal reasoning, and knowledge update, LiCoMemory achieves 73.80% accuracy and 76.63% recall with GPT 4o mini. These gains show that LiCoMemory retrieves more relevant long term context while using fewer tokens and lower latency than Mem0g and MemOS.
BENCHMARK
On LongMemEval, which tests single session, multi session, temporal reasoning, and knowledge update, LiCoMemory achieves 73.80% accuracy and 76.63% recall with GPT 4o mini. These gains show that LiCoMemory retrieves more relevant long term context while using fewer tokens and lower latency than Mem0g and MemOS.
BENCHMARK
Acc. on LongMemEval with GPT 4o mini.
BENCHMARK
Accuracy on LoCoMo in real time interaction setting.
KEY INSIGHT
LiCoMemory cuts construction latency on LoCoMo from 1772s in Mem0 to 21s while increasing accuracy from 54.68% to 66.4%.
This is surprising because graph based memory is usually assumed to be slower than vector memory, yet CogniGraph shows a 84x speedup with better accuracy.
WHY IT MATTERS
LiCoMemory enables agents to maintain hierarchical, temporally aware memories that update in real time without heavy graph reconstruction.
Builders can now deploy long running assistants that answer temporal and multi session questions accurately while keeping token budgets and latency low enough for interactive use.
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%.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…