AuthorsHan Chen, Zining Zhang, Wenqi Pei et al.
TL;DR
MemForest uses hierarchical temporal MemTrees with parallel extraction and dirty-path refresh to reach 81.8% pass@1 on LongMemEval-S while building memory 6.0× faster than EverMemOS.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Many agent memory systems keep autoregressive extraction and state-dependent maintenance on the write path, creating long critical paths before new dialogue becomes queryable.
On LongMemEval-S and LoCoMo, this design leaves systems like EverMemOS with much lower input-normalized build rates, causing stale memory and delayed access to new evidence for long-horizon reasoning.
HOW IT WORKS
MemForest’s core mechanism combines parallel extraction, a shared memory substrate of canonical facts, scoped MemTrees, and forest recall plus tree browse for temporal agent memory.
You can think of MemForest like a database-backed card catalog: canonical facts are the cards, and MemTrees are time-ordered drawers that can be updated and searched without rewriting the whole catalog.
This KEY_MECHANISM of scoped temporal trees with dirty-path refresh lets MemForest preserve historical trajectories and transitions that a plain context window or flat vector store cannot represent or retrieve reliably.
DIAGRAM
This diagram shows how MemForest answers a user query by recalling relevant MemTrees and browsing from interval summaries down to leaf evidence.
DIAGRAM
This diagram shows how MemForest is evaluated: building persistent memory from dialogue streams, then answering benchmark questions with shared backbones and a fixed judge.
PROCESS
Parallel Extraction
MemForest partitions each session into short chunks and runs parallel extraction, producing local memory candidates that avoid a single serialized LLM pass.
Canonical Fact Manager
MemForest canonicalizes chunk outputs into ManagedFacts, merging duplicates and normalizing entities so later updates can reuse stable write units.
Entity and Scene Router
MemForest routes canonical facts into session, entity, and scene scopes, emitting update records that determine which MemTrees each fact should materialize in.
MemTree Dirty Refresh
MemForest inserts leaves into MemTrees, marks ancestor paths dirty, and refreshes summaries and embeddings level by level, updating RootIndex and NodeIndex for retrieval.
KEY CONTRIBUTIONS
Write-efficient temporal agent memory formulation
MemForest frames continuous agent memory as a write-efficient temporal data-management problem, identifying serialized extraction and state-dependent maintenance as key freshness bottlenecks (Table 1).
MemForest architecture with MemTree hierarchical temporal index
MemForest introduces MemTrees that organize each temporal scope as a time-ordered tree, enabling localized dirty-path refresh and coarse-to-fine retrieval over canonical facts.
Improved speed–accuracy trade-off on LongMemEval-S and LoCoMo
MemForest reaches 81.8% pass@1 on LongMemEval-S and 84.09% on LoCoMo categories 1–4 with Qwen3-30B, while achieving 6.0× and 9.5× higher build rates than EverMemOS on matched traces.
RESULTS
LongMemEval-S overall pass@1
81.80%
+14.80 over EverMemOS
LoCoMo categories 1–4 overall
84.09%
-0.13 vs EverMemOS
LongMemEval-S build rate (turns/s)
2.841
6.0× EverMemOS build rate of 0.471
LoCoMo build rate (turns/s)
7.020
9.5× EverMemOS build rate of 0.738
These metrics come from LongMemEval-S and LoCoMo, which stress single-session, multi-session, and temporal reasoning for long-lived agents.
The MAIN_RESULT shows that MemForest can maintain high answer quality while drastically reducing memory build latency compared to EverMemOS and other baselines.
BENCHMARK
These metrics come from LongMemEval-S and LoCoMo, which stress single-session, multi-session, and temporal reasoning for long-lived agents. The MAIN_RESULT shows that MemForest can maintain high answer quality while drastically reducing memory build latency compared to EverMemOS and other baselines.
BENCHMARK
Pass@1 accuracy (%) on LongMemEval-S using the public benchmark prompt and a fixed judge.
BENCHMARK
Pass@1 accuracy (%) on LoCoMo categories 1–4 under the official mapping.
KEY INSIGHT
MemForest achieves 84.09% pass@1 on LoCoMo categories 1–4 with Qwen3-30B, only 0.13 points below EverMemOS despite a 9.5× faster build rate.
This is counterintuitive because we usually expect aggressive write-path optimization to hurt accuracy, yet MemForest preserves temporal fidelity well enough to match a slower, more heavyweight memory system.
WHY IT MATTERS
MemForest unlocks persistent agent memory that can ingest long-horizon interaction streams quickly while still answering temporal and multi-session questions accurately.
Builders can now deploy agents that maintain rich, temporally structured memories over weeks or months without incurring prohibitive write latency or needing to replay entire histories for maintenance.
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…