AuthorsZeyang Cui, Jiannong Cao, Zhiyuan Wen et al.
TL;DR
EdgeMem uses a multi-anchor hypergraph over original dialogue turns to deliver 61.01 strict-judge accuracy on LoCoMo while using 0 generative-LLM tokens for memory.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing agent memories compress histories with LLM-written summaries, where memory construction alone can consume 11.77M tokens versus a 174.5k-token raw history.
When LoCoMo-scale agents like A-Mem or MemoryOS summarize repeatedly, answer-bearing details vanish and LLM-based retrieval later returns wrong answers despite large token spend.
HOW IT WORKS
EdgeMem’s core mechanism keeps every turn verbatim and builds a Multi-anchor Hypergraph Construction with Time Sub-Hypergraph, Co-occurrence Sub-Hypergraph, and Episode Sub-Hypergraph plus Memory Retrieval and a Generated Answer reader.
Think of EdgeMem like a library card catalog: each conversation turn is a book, and the three sub-hypergraphs are parallel catalogs by topic, date, and episode shelf.
This multi-anchor design lets EdgeMem recall precise evidence across sessions and time, something a plain context window or single BM25 index cannot match.
DIAGRAM
This diagram shows how EdgeMem routes a user query through cue extraction and the three Sub-Hypergraphs to build the final evidence pack.
DIAGRAM
This diagram shows how EdgeMem is evaluated on LoCoMo, including anchor ablations and shared answer prompting against baselines.
PROCESS
Local Annotation for Anchor Keys
EdgeMem runs Local Annotation for Anchor Keys to extract entities, keywords, and dates, attaching them to each original turn without any LLM calls.
Multi-anchor Hypergraph Construction
EdgeMem performs Multi-anchor Hypergraph Construction, linking turns into Time Sub-Hypergraph, Co-occurrence Sub-Hypergraph, and Episode Sub-Hypergraph using the annotated keys.
Memory Retrieval
During Memory Retrieval, EdgeMem computes BM25 scores, applies the Episode channel, Co-occurrence channel, and Time channel, and fuses their outputs under a fixed evidence budget.
Generated Answer
In the Generated Answer stage, EdgeMem sends the query and serialized evidence pack to a single reader call, which returns either a grounded answer or the refusal Not mentioned in the conversation.
KEY CONTRIBUTIONS
Source-preserving multi-anchor hypergraph memory
EdgeMem introduces a source-preserving Multi-anchor Hypergraph Construction where Time Sub-Hypergraph, Co-occurrence Sub-Hypergraph, and Episode Sub-Hypergraph enable LLM-free memory construction and retrieval.
Deterministic budget-aware retrieval policy
EdgeMem designs a deterministic Memory Retrieval policy that prioritizes the Episode Sub-Hypergraph as the evidence core and additively incorporates Co-occurrence and Time Sub-Hypergraph evidence under a fixed budget.
Efficient high-quality agent memory on LoCoMo and LongMemEval-S
EdgeMem achieves 61.01 strict-judge accuracy on LoCoMo and 81.49 Recall@3 on LongMemEval-S while using 0 construction tokens and only 1,000 tokens per question.
RESULTS
SJ
61.01
+2.31 over CompassMem
Tokens per question
1,000 tokens
-40,348 vs CompassMem
Recall@3
81.49
+2.98 over MemGAS on LongMemEval-S
Build tokens
0 tokens
-11.77M vs A-Mem on LoCoMo
On LoCoMo, which tests long-horizon conversational question answering, EdgeMem reaches 61.01 strict-judge accuracy versus 58.70 for CompassMem while using only 1,000 tokens per question. On LongMemEval-S, which evaluates session-level retrieval over roughly 115k-token haystacks, EdgeMem attains 81.49 Recall@3 and 90.49 NDCG@3 with 3,202 tokens per question.
BENCHMARK
On LoCoMo, which tests long-horizon conversational question answering, EdgeMem reaches 61.01 strict-judge accuracy versus 58.70 for CompassMem while using only 1,000 tokens per question. On LongMemEval-S, which evaluates session-level retrieval over roughly 115k-token haystacks, EdgeMem attains 81.49 Recall@3 and 90.49 NDCG@3 with 3,202 tokens per question.
BENCHMARK
Strict-judge (SJ) accuracy on LoCoMo under the shared answer prompt, comparing EdgeMem against leading memory systems.
BENCHMARK
Total generative-LLM build tokens on LoCoMo, highlighting EdgeMem’s LLM-free construction compared to token-heavy baselines.
KEY INSIGHT
EdgeMem reaches 61.01 strict-judge accuracy on LoCoMo while using only 1,000 tokens per question and 0 construction tokens.
This is surprising because systems like CompassMem spend 41,348 tokens per question and 1.76M build tokens, yet still trail EdgeMem by 2.31 strict-judge points.
WHY IT MATTERS
EdgeMem shows that preserving original turns and organizing them with multi-anchor hypergraphs can beat LLM-written memories in both accuracy and abstention.
Builders can now deploy long-horizon agents with faithful, low-cost memory that scales without generative maintenance, making multi-session reasoning practical even under strict token budgets.
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…