AuthorsYaxiong Wu, Yongyue Zhang, Sheng Liang, Yong Liu
TL;DR
SGMem uses sentence graph memory over chunked dialogue to align raw history with generated memory, reaching 0.700 Accuracy on LongMemEval vs 0.676 for RAG-SMFI.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Long term conversational agents face memory overload, where stored content exceeds their ability to manage and retrieve it effectively, causing fragmented context.
Existing RAG based QA systems scatter information across raw dialogue and generated memory, leading to memory fragmentation that undermines temporal reasoning and coherent, user tailored responses.
HOW IT WORKS
SGMem centers on SGMem Construction and Management and SGMem Usage, building sentence level graphs that connect sessions, rounds, turns, sentences, summaries, facts, and insights.
You can think of SGMem as a conversational card catalog, where each sentence is an index card linked to related cards and higher level folders like sessions and rounds.
This sentence graph memory lets SGMem traverse multi hop associations and align raw dialogue with generated memory, providing coherent context that a plain context window cannot assemble.
DIAGRAM
This diagram shows how SGMem Usage retrieves heterogeneous memories, expands sentences via graph traversal, ranks chunks, and feeds context to the LLM for QA.
DIAGRAM
This diagram shows how SGMem is evaluated on LongMemEval and LoCoMo with different RAG and SGMem variants and hyperparameter ablations.
PROCESS
Processing Conversations
SGMem decomposes sessions into rounds, turns, and sentences, and uses an LLM to generate summaries, facts, and insights that complement the raw dialogue.
Indexing
SGMem embeds sessions, rounds, turns, sentences, summaries, facts, and insights with Sentence BERT, building seven searchable index tables in a vector database.
Constructing Sentence Graph Memory
SGMem links chunk nodes to their sentences and connects sentences via k nearest neighbor similarity edges, forming a sentence graph stored in a graph database.
SGMem Usage
SGMem retrieves heterogeneous memories, expands sentences via h hop traversal, ranks parent chunks, aggregates Crelevant, and conditions the LLM to generate the final answer.
KEY CONTRIBUTIONS
Sentence Graph Memory Construction and Management
SGMem introduces hierarchical sentence graph memory that links chunk nodes and sentences via similarity edges, mitigating memory fragmentation across turns, rounds, and sessions.
Sentence Graph Memory Usage
SGMem designs a multi hop retrieval mechanism over sentence graphs that integrates raw dialogue with summaries, facts, and insights into a coherent Crelevant for QA.
Comprehensive Evaluation
SGMem is extensively evaluated on LongMemEval and LoCoMo, achieving up to 0.700 Accuracy on LongMemEval and 0.526 on LoCoMo, surpassing strong memory and RAG baselines.
RESULTS
Accuracy Top 5 LongMemEval
0.700
+0.024 over RAG-SMFI
Accuracy Top 10 LongMemEval
0.730
+0.050 over RAG-SMFI
Accuracy Top 5 LoCoMo
0.526
+0.016 over RAG-SMFI
Accuracy Top 10 LoCoMo
0.532
+0.004 over RAG-SMFI
Table 1 reports Accuracy on LongMemEval and LoCoMo, which test long term conversational QA with multi session, temporal, and knowledge update queries. The gains show that SGMem leverages sentence graph memory to retrieve more coherent context than RAG-SMFI despite using the same underlying LLM.
BENCHMARK
Table 1 reports Accuracy on LongMemEval and LoCoMo, which test long term conversational QA with multi session, temporal, and knowledge update queries. The gains show that SGMem leverages sentence graph memory to retrieve more coherent context than RAG-SMFI despite using the same underlying LLM.
BENCHMARK
Accuracy Top 5 on LongMemEval for SGMem and strong RAG baselines.
KEY INSIGHT
On LoCoMo, SGMem SF reaches 0.522 Accuracy Top 5, slightly above SGMem SMFI at 0.526, despite using fewer memory types than full SMFI.
This is surprising because adding more summaries, facts, and insights is expected to always help, but SGMem shows that excessive generated memory can introduce noise in very long conversations.
WHY IT MATTERS
SGMem unlocks sentence level graph memory that aligns raw dialogue and generated memory without expensive entity extraction, enabling coherent retrieval across long, multi session histories.
Builders can now deploy long term conversational agents that scale beyond context windows while preserving fine grained temporal and personalization cues that were previously lost or fragmented.
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.
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.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…