AuthorsKaixiang Wang, Yidan Lin, Jiong Lou et al.
TL;DR
E-mem uses multi-agent episodic context reconstruction to keep full raw histories, reaching 54.17% F1 on LoCoMo (+8.86 over GAM) with ~43× lower large-model token cost.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory preprocessing compresses rich sequential histories into static embeddings or graphs, causing destructive de contextualization that severs critical causal chains for deep reasoning.
This failure shows up on information dense benchmarks like LoCoMo, where traditional memory systems struggle to reconstruct complex multi hop and temporal dependencies from fragmented chunks.
HOW IT WORKS
E-mem centers on Master Agent, Assistant Agents, a Multi-Pathway Routing Mechanism, Sliding Window Segmentation with Overlap, and Episodic Memory Context Retention and Isolation to keep uncompressed episodic segments.
Think of E-mem like a brain with a hippocampus: assistants hold detailed episodic memories, while the master agent behaves like a prefrontal cortex orchestrating high level reasoning.
This architecture lets E-mem re experience past events inside original token sequences, enabling temporally grounded evidence extraction that a plain context window or static RAG index cannot provide.
DIAGRAM
This diagram shows how E-mem routes a query, activates assistants, reconstructs episodic contexts, and aggregates temporally anchored evidence into a final answer.
DIAGRAM
This diagram shows how E-mem is evaluated on LoCoMo and HotpotQA, including routing pathway ablations and chunk activation settings.
PROCESS
Memory Building and Storing
E-mem uses Sliding Window Segmentation with Overlap to form episodic contexts Ei and Episodic Memory Context Retention and Isolation to assign each Ei to an Assistant Agent.
Memory Activation via Multi Pathway Routing
The Multi-Pathway Routing Mechanism computes Pglobal, Pvec, and Pkw over summaries si and contexts Ei, selecting a set of Assistant Agents A* for activation.
Episodic Context Reconstruction and Local Reasoning
Activated Assistant Agents re experience their preserved Ei and output temporally anchored evidence tuples ei = <ci, τi> via the local mapping Φasst.
Synergistic Reasoning and Response Generation
The Master Agent aggregates evidence set E, enforces chronological consistency with Ψmaster, and optionally runs iterative Refine and Query steps before returning the final response.
KEY CONTRIBUTIONS
Episodic Context Reconstruction
E-mem delegates local reasoning to Assistant Agents that preserve full Ei segments, surfacing only logically deduced evidence and achieving 54.17 F1 on LoCoMo overall.
Heterogeneous Hierarchical Master Assistant Architecture
E-mem decouples high level planning in the Master Agent from storage in SLM based Assistant Agents, avoiding lost in the middle effects across long horizons.
SOTA Performance with Token Efficiency
E-mem’s Multi-Pathway Routing Mechanism and episodic reconstruction deliver +8.56 F1 on multi hop and +8.87 F1 on temporal LoCoMo subsets while reducing token cost by over 70%.
RESULTS
LoCoMo Overall F1
54.17
+8.86 over GAM with GPT-4o-mini
LoCoMo BLEU-1
44.34
+6.56 over GAM with GPT-4o-mini
HotpotQA F1 1600 docs
55.76
+2.05 over GAM with GPT-4o-mini
Normalized Cost Tokens
3621
43× lower than Long-Context 169100 on LoCoMo
On LoCoMo, which tests single hop, multi hop, temporal, and open domain long term memory, E-mem reaches 54.17 F1 versus 45.31 for GAM using GPT-4o-mini. On HotpotQA with 1600 documents, E-mem attains 55.76 F1 while keeping normalized large model token cost to 3621, demonstrating efficient episodic reconstruction at ultra long scales.
BENCHMARK
On LoCoMo, which tests single hop, multi hop, temporal, and open domain long term memory, E-mem reaches 54.17 F1 versus 45.31 for GAM using GPT-4o-mini. On HotpotQA with 1600 documents, E-mem attains 55.76 F1 while keeping normalized large model token cost to 3621, demonstrating efficient episodic reconstruction at ultra long scales.
BENCHMARK
Overall F1 on LoCoMo for E-mem and leading baselines using GPT-4o-mini as master agent.
BENCHMARK
F1 on HotpotQA 1600 document setting for E-mem and baselines with GPT-4o-mini.
KEY INSIGHT
E-mem achieves near optimal reasoning with only 8 activated chunks on HotpotQA, reaching 55.76 F1, while more chunks yield only marginal gains.
This is surprising because typical intuition suggests feeding more context should help, yet E-mem’s router shows that carefully selected episodic segments beat brute force concatenation.
WHY IT MATTERS
E-mem unlocks scalable System 2 style reasoning where agents can re experience long histories without compressing them into lossy embeddings or graphs.
Builders can now design agents that maintain logically consistent multi session memories, perform temporally grounded multi hop reasoning, and keep token costs manageable using heterogeneous master assistant deployments.
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…