AuthorsYunxiang Zhang, Yiheng Li, Ali Payani, Lu Wang
TL;DR
AdaMEM uses a hybrid long-term trajectory memory plus short-term strategy memory with STEP-MFT, reaching 58.2% unseen ALFWorld success (+6.0 over Synapse).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
AdaMEM targets agents that only retrieve memory once at episode start, forcing static guidance that drifts on long-horizon tasks and hurts adaptation.
On ALFWorld and WebShop, this rigidity leaves static memories up to 13% and 11% behind AdaMEM, causing agents to repeat failing behaviors instead of correcting mid-trajectory.
HOW IT WORKS
AdaMEM’s core mechanism is a hybrid of Long-Term Trajectory Memory, Short-Term Strategy Memory, ADAMEM-HIGH, ADAMEM-LOW, and STEP-MFT for step-wise test-time adaptation.
You can think of Long-Term Trajectory Memory as disk storing rich past trajectories, while Short-Term Strategy Memory acts like RAM holding a compact, situation-specific plan.
This design lets AdaMEM synthesize fresh strategies conditioned on the current state, enabling continuous adaptation that a fixed context window or static prompt cannot provide.
DIAGRAM
This diagram shows how AdaMEM-HIGH and AdaMEM-LOW decide when to retrieve from Long-Term Trajectory Memory and update Short-Term Strategy Memory at each step.
DIAGRAM
This diagram shows how AdaMEM collects decision tuples with ADAMEM-HIGH, filters them with STEP-MFT, and evaluates on ALFWorld and WebShop.
PROCESS
Long-Term Trajectory Memory Construction
AdaMEM first builds Long-Term Trajectory Memory M offline from successful trajectories, storing state embeddings and future action sequences for dense retrieval.
Short-Term Strategy Memory Synthesis
At test time, AdaMEM retrieves experiences E_ret and generates Short-Term Strategy Memory z_t conditioned on the current state to guide immediate decisions.
ADAMEM-HIGH and ADAMEM-LOW Inference
Using ADAMEM-HIGH or ADAMEM-LOW, AdaMEM decides when to refresh strategies, trading off token cost against adaptability during the episode.
STEP-MFT Strategy Fine-Tuning
With STEP-MFT, AdaMEM filters successful, action-changing strategies and performs supervised fine-tuning so future strategies more reliably alter actions toward success.
KEY CONTRIBUTIONS
Adaptive Memory Agent (ADAMEM)
AdaMEM introduces a hybrid memory framework combining Long-Term Trajectory Memory and Short-Term Strategy Memory to enable continuous test-time adaptation without parameter updates, yielding up to 13% relative gains on ALFWorld.
STEP-MFT for Strategy Learning
AdaMEM proposes STEP-MFT, a step-wise Memory Fine-Tuning technique that filters strategies by outcome success and action change to train high-utility strategy generation.
Scaling Dimension for Agentic Memory
AdaMEM demonstrates a new scaling dimension where increasing retrieval budget k and adaptation effort improves performance, achieving up to 17% relative gains on ALFWorld and 13% on WebShop over static memory baselines.
RESULTS
ALFWorld Seen Success Rate (%)
54.0%
+1.9 over Synapse
ALFWorld Unseen Success Rate (%)
58.2%
+6.0 over Synapse
WebShop Task Score
74.2
+2.8 over No Memory
HotpotQA Success Rate (%)
41.1%
+0.7 over ReasoningBank
On ALFWorld and WebShop, which test embodied navigation and web shopping, AdaMEM’s dynamic memory raises unseen ALFWorld success to 58.2% and WebShop Task Score to 74.2, surpassing static memories like Synapse and ReasoningBank. On HotpotQA cross-episode search, AdaMEM reaches 41.1% Success Rate, slightly ahead of both static memory baselines.
BENCHMARK
On ALFWorld and WebShop, which test embodied navigation and web shopping, AdaMEM’s dynamic memory raises unseen ALFWorld success to 58.2% and WebShop Task Score to 74.2, surpassing static memories like Synapse and ReasoningBank. On HotpotQA cross-episode search, AdaMEM reaches 41.1% Success Rate, slightly ahead of both static memory baselines.
BENCHMARK
Success Rate (%) on ALFWorld unseen for different memory mechanisms with Qwen3-4B-Instruct.
KEY INSIGHT
On WebShop, static memories hurt performance: Synapse drops Task Score from 71.4 with No Memory down to 65.4, a 6.0 point decrease.
This is surprising because more experience usually helps, but retrieving full trajectories at episode start from an uninformative homepage injects noisy priors that mislead the agent.
WHY IT MATTERS
AdaMEM unlocks agents that can continuously adapt strategies at test time using external experience, without any online parameter updates or extra critics.
Builders can now scale memory banks and test-time compute to trade tokens for robustness, enabling self-evolving agents in navigation, web interaction, and multi-hop search settings.
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…