AuthorsYongxian Wei, Yilin Zhao, Runxi Cheng et al.
TL;DR
MemAgent uses content aware routing over heterogeneous memory providers to raise average accuracy to 75.7% on GAIA WebWalkerQA and xBench DS, +10.0 points over No Memory.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
MemAgent’s study of 13 memory methods shows that no single provider dominates, while an oracle over all providers reaches 89.6% accuracy versus 71.0% for the best single provider.
This means agents with a fixed memory representation waste complementary strengths across tasks, leaving up to 18.6 percentage points of accuracy unrealized and limiting long horizon and cross task performance.
Without heterogeneous routing, agents remain largely stateless across tasks, preventing continual improvement and amortization of effort on GAIA, WebWalkerQA, and xBench DS.
HOW IT WORKS
MemAgent combines a content aware routing architecture, a training data synthesis pipeline, InsightGraph, and a short term memory provider to manage BEGIN, IN, and END memory decisions.
You can think of MemAgent as a smart operating system that chooses between different memory disks and a RAM buffer, instead of relying on one oversized but clumsy drive.
This KEY_MECHANISM lets MemAgent preview, gate, and selectively store experiences so agents reuse the right memories at the right time, beyond what a plain context window can encode.
DIAGRAM
This diagram shows how MemAgent routes memory at BEGIN, IN, and END phases between the task agent and heterogeneous providers.
DIAGRAM
This diagram shows how MemAgent synthesizes supervision via forced exploration, per decision rewards, and reward filtered supervised fine tuning.
PROCESS
BEGIN phase content aware probing
MemAgent uses the content aware routing architecture to probe all providers, including InsightGraph and other long term memory providers, before selecting one for retrieval.
IN phase short term memory gating
During each step, MemAgent consults the short term memory provider and decides whether to inject a working memory summary or skip injection.
END phase multi provider storage
After the trajectory finishes, MemAgent uses the content aware routing architecture to select a subset of providers for storing the trajectory for future reuse.
Training data synthesis pipeline
Offline, MemAgent’s training data synthesis pipeline runs forced exploration, per decision reward scoring, and reward filtered supervised fine tuning to learn routing policies.
KEY CONTRIBUTIONS
Rethinking agent memory with heterogeneous providers
MemAgent unifies 13 memory methods as providers, adds InsightGraph and a short term memory provider, and shows an oracle reaches 89.6% versus 71.0% for the best single provider.
Three phase routing formulation BEGIN IN END
MemAgent formulates memory as routing decisions at BEGIN, IN, and END, implemented by a content aware routing architecture that previews, gates, and selectively stores experiences.
Training data synthesis for routing without labels
MemAgent introduces a training data synthesis pipeline combining forced exploration, LLM as Judge rewards, and downstream reuse signals to supervise routing without ground truth labels.
RESULTS
GAIA Overall
69.7%
+10.9 over No Memory
WebWalkerQA
79.4%
+10.0 over No Memory
xBench DS
78.0%
+9.0 over No Memory
Avg.
75.7%
+10.0 over No Memory
On GAIA, WebWalkerQA, and xBench DS, which test multi step tool use and deep search, MemAgent raises average accuracy from 65.7% to 75.7%, proving that heterogeneous routing yields a 10.0 percentage point gain over a stateless No Memory agent.
BENCHMARK
On GAIA, WebWalkerQA, and xBench DS, which test multi step tool use and deep search, MemAgent raises average accuracy from 65.7% to 75.7%, proving that heterogeneous routing yields a 10.0 percentage point gain over a stateless No Memory agent.
BENCHMARK
Accuracy (%) of MemAgent versus No Memory, best single provider, and PromptRoute across three benchmarks.
KEY INSIGHT
MemAgent shows that an oracle over 13 providers reaches 89.6% accuracy, while the best individual provider only reaches 71.0%.
This breaks the assumption that designing a single superior memory representation is enough, revealing that routing across heterogeneous memories is crucial.
Even sophisticated single providers like AWM and ExpeL lead on only one benchmark each, contradicting expectations of a universal memory format.
WHY IT MATTERS
MemAgent unlocks adaptive, content aware orchestration of diverse memory systems, raising average accuracy by 10.0 percentage points while reducing task steps by 12%.
Builders can now plug multiple specialized memory providers into one agent and let MemAgent learn when to retrieve, inject, and store, instead of hand picking a single memory design.
Jingbiao Mei, Jinghong Chen et al.
arXiv 2026 · 2026
ATM-Bench structurally evaluates long-term multimodal personal memory using Memory Ingestion, Retrieval, and Answer Generation with Schema-Guided Memory and Descriptive Memory variants. On ATM-Bench-Hard, Oracle with SGM reaches 47.3% QS while the best full system stays under 20% accuracy, revealing a large gap.
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.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…