AuthorsJunxi Wang, Te Sun, Jiayi Zhu et al.
TL;DR
MemForest uses EventTree Semantic-Temporal Partitioning plus EventTree Progressive Merging to compress 50% of agent memory while retaining 97.1% performance and achieving up to 2.24× retrieval speedup.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Agent memory systems accumulate historical memory over long interactions, and Mem0 shows that compressing 50% still needs to preserve 97.1% performance while achieving 1.89× speedup.
As memory nodes grow, embedding time, retrieval time, and response time all increase, making frameworks like Mem0 and M3-Agent struggle with storage pressure and slower retrieval.
HOW IT WORKS
MemForest centers on EventTree Semantic-Temporal Partitioning, EventTree Progressive Merging, and Anchor-Guided Propagation Retrieval, which together restructure and compress historical memory into event-centric trees.
You can think of MemForest as turning a messy log of memories into a forest of event trees, like a card catalog that groups related cards by both topic and time.
By merging highly similar nodes inside each EventTree and then using anchor-guided temporal propagation, MemForest preserves event structure and retrieval accuracy far beyond what a plain context window or naive pruning can keep.
DIAGRAM
This diagram shows how MemForest performs Anchor-Guided Propagation Retrieval from a compressed memory set for a single query.
DIAGRAM
This diagram shows how MemForest is evaluated on Mem0 and M3-Agent with different compression ratios and baselines.
PROCESS
EventTree Semantic-Temporal Partitioning
MemForest first applies EventTree Semantic-Temporal Partitioning using KMeans clustering with global similarity gj_ci and local continuity lj_ci to form event units Ck.
EventTree Progressive Merging
MemForest builds a maximum spanning tree Tk per event unit and progressively merges node pairs using updated weights w'_ij until the target compression ratio is reached.
Anchor-Guided Propagation Retrieval
MemForest retrieves a candidate set Mh, selects anchor nodes Ma, computes propagation scores pj, and fuses them with similarity scores sj into vj.
Query Answering with Compressed Memory
MemForest feeds the top k memories selected by fusion scores vj into the agent backbone, enabling Mem0 or M3-Agent to answer while using only 50% of the original memory.
KEY CONTRIBUTIONS
General Compression Framework
MemForest introduces a general agent memory compression framework that combines EventTree Semantic-Temporal Partitioning and EventTree Progressive Merging to compress 50% of historical memory while retaining 97.1% and 99.7% performance on Mem0 and M3-Agent.
Novel Retrieval Mechanism
MemForest adds Anchor-Guided Propagation Retrieval, which fuses query similarity and temporal propagation scores to recover relevant neighbors and yields up to 21.0% performance gains after 50% compression in M3-Agent.
Excellent Empirical Performance
MemForest achieves 1.89× retrieval speedup on Mem0 and 2.24× on M3-Agent at 50% compression, outperforming Random Pruning, KMeans, DART, StreamMeCo, Random Merging, and ToMe in performance retention.
RESULTS
Avg. All Mem0
59.3 score
+6.3 over Random Pruning at 50% on Mem0
Avg. All M3Agent
39.0 score
+4.3 over Random Pruning at 50% on M3-Agent
Retention Mem0
97.1%
MemForest vs Mem0 baseline at 50% compression
Retention M3Agent
99.7%
MemForest vs M3-Agent baseline at 50% compression
On LoCoMo, LongMemEval, PersonaMem, M3-Bench-robot, and M3-Bench-web, MemForest compresses 50% of historical memory while keeping 59.3 vs 61.1 average score on Mem0 and 39.0 vs 39.1 on M3-Agent, proving MemForest preserves nearly all accuracy with large retrieval speedups.
BENCHMARK
On LoCoMo, LongMemEval, PersonaMem, M3-Bench-robot, and M3-Bench-web, MemForest compresses 50% of historical memory while keeping 59.3 vs 61.1 average score on Mem0 and 39.0 vs 39.1 on M3-Agent, proving MemForest preserves nearly all accuracy with large retrieval speedups.
BENCHMARK
Avg. All score on Mem0 with 50% historical memory compression.
BENCHMARK
Avg. All score on M3-Agent with 50% historical memory compression.
KEY INSIGHT
MemForest retains 99.7% of M3-Agent performance at 50% compression, with Avg. All score 39.0 vs 39.1 while achieving a 2.24× retrieval speedup.
This is surprising because aggressive memory compression is usually expected to hurt multimodal agents more, yet MemForest shows multimodal memory has enough redundancy to compress by half with almost no accuracy loss.
WHY IT MATTERS
MemForest unlocks event-level memory compression that respects both global semantic similarity and local temporal continuity, enabling long-running agents to keep memory growth under control.
Builders can now deploy Mem0 and M3-Agent style systems that maintain near-original accuracy while halving memory size and nearly doubling retrieval speed, making persistent agents practical at scale.
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…