AuthorsYunhan Jiang, Wenbin Duan, Shasha Guo et al.
TL;DR
ActiveMem uses a decoupled Planner plus Distributed Memory System to store query-conditioned gists and reaches 0.79 LasJ on BrowseComp-Plus with only 2145 PFLOPs.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
ActiveMem targets centralized memory systems where context compression causes permanent information loss, leaving memory content irreversibly degraded and unrecoverable for later reasoning.
In long-horizon web reasoning like BrowseComp-Plus and GAIA, this forces agents to choose between context overload and losing evidence, degrading downstream LasJ accuracy as trajectories grow.
HOW IT WORKS
ActiveMem’s core mechanism separates a high-level Planner from a Distributed Memory System composed of Memorizers, Memory Shards, and an Operator that manage query-conditioned gists.
Like a prefrontal cortex issuing control signals to a hippocampus, the Planner sends focused queries while Memorizers and Memory Shards act as distributed long-term storage.
This decoupling lets ActiveMem keep the Planner’s context clean and compact while preserving document-level information in shards, avoiding the centralized trade-off between context length and memory fidelity.
DIAGRAM
This diagram shows how ActiveMem routes Planner queries through the Operator and Memorizers to Memory Shards and back as distilled gists at inference time.
DIAGRAM
This diagram shows how ActiveMem is evaluated on BrowseComp-Plus and GAIA and how ablations on Memory Shards and Memorizer variants are organized.
PROCESS
Planner
The Planner maintains state s_t = (x, h_t, m_{t-1}), trims history to the most recent K steps, and emits retrieval queries Q_t.
Distributed Memory System
The Distributed Memory System receives retrieved documents D_t, routes them to Memory Shards, and orchestrates Memorizers and the Operator for gist distillation.
Memorizers
Memorizers take each (q, c) pair, generate query-conditioned gists g_c = ω(c, q), and write them into the appropriate Memory Shards.
Operator
The Operator detects memory hits or misses, reuses or updates shard entries, consolidates gists asynchronously, and returns distilled gists G_t to the Planner.
KEY CONTRIBUTIONS
Neuroscience-inspired cognitive decoupling
ActiveMem draws on prefrontal cortex and hippocampus complementarity to separate the Planner’s executive control from distributed memory consolidation in Memorizers and Memory Shards.
The ActiveMem framework
ActiveMem introduces a Distributed Memory System where parallel Memorizers distill gists, Memory Shards persist them, and an Operator coordinates routing and consolidation.
Superior accuracy with lower computational cost
On BrowseComp-Plus, ActiveMem reaches 0.79 LasJ with 2145 PFLOPs, improving MemoBrain’s 0.63 LasJ at 1640 PFLOPs by +0.16 LasJ.
RESULTS
LasJ
0.79
+0.07 over Context-Folding
PFLOPs (BrowseComp-Plus)
2145
2145 vs 3920 for Context-Folding
LasJ (GAIA)
0.62
+0.03 over AgentFold and MemoBrain
PFLOPs (GAIA)
187
187 vs 516 for MemoBrain
On BrowseComp-Plus and GAIA WebSearch, which test long-horizon web reasoning, ActiveMem delivers higher LasJ accuracy at substantially lower PFLOPs than centralized memory agents and vanilla ReAct LLMs.
BENCHMARK
On BrowseComp-Plus and GAIA WebSearch, which test long-horizon web reasoning, ActiveMem delivers higher LasJ accuracy at substantially lower PFLOPs than centralized memory agents and vanilla ReAct LLMs.
BENCHMARK
Total LasJ score on BrowseComp-Plus across Easy, Medium, and Hard splits.
BENCHMARK
PFLOPs per module in ActiveMem on BrowseComp-Plus.
KEY INSIGHT
ActiveMem processes 188M tokens on BrowseComp-Plus, more than centralized agents’ 47–167M, yet uses only 2145 PFLOPs and still reaches 0.79 LasJ.
This is surprising because more tokens usually imply higher cost, but ActiveMem’s small 4B Memorizers handle most tokens cheaply while the large Planner stays context-light.
WHY IT MATTERS
ActiveMem shows that decoupling reasoning from memory lets agents scale to long-horizon tasks while keeping a compact, high-quality reasoning context.
Builders can now design agents that aggressively read and reuse web evidence via distributed gists, without being constrained by a single monolithic context window.
Mustafa Arslan
· 2026
Aeon restructures LLM memory using the Atlas, Trace, Semantic Lookaside Buffer, Write Ahead Log, and Sidecar Blob Arena inside a zero copy Core Shell kernel. Aeon achieves 4.70 ns INT8 dot products, 3.09 µs Atlas traversal at 100K nodes, 3.1× compression, and P99 read latency of 750 ns under 16 thread contention compared to FP32 and flat scan baselines.
Bin Wen, Ruoxuan Zhang et al.
· 2026
Neuro-Symbolic Dual Memory Framework uses Progress Memory, Feasibility Memory, a Blueprint Planner Agent, a Progress Monitor Agent, and an Actor Agent to decouple semantic progress guidance from executable feasibility checks. On ALFWorld, Neuro-Symbolic Dual Memory Framework achieves 94.78% success rate versus 88.81% for AWM, and on WebShop reaches 0.7132 score versus 0.5998 for WALL-E 2.0.
Zhixing You, Jiachen Yuan, Jason Cai
· 2026
D-Mem combines Mem0∗, Quality Gating, and Full Deliberation into a dual-process memory system that incrementally stores vector memories and selectively scans raw history. On LoCoMo with GPT-4o-mini, D-Mem’s Quality Gating reaches 53.5 F1 versus the Mem0∗ baseline’s 51.2 F1, recovering 96.7% of the 55.3 F1 Full Deliberation performance with far fewer tokens.
Answers use this explainer on Memory Papers.
Checking…