AuthorsWei-Chieh Huang, Weizhi Zhang, Yueqing Liang et al.
TL;DR
Rethinking Memory Mechanisms of Foundation Agents in the Second Half unifies internal–external substrates, five cognitive mechanisms, and user vs agent subjects into a single taxonomy over 218 papers.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Rethinking Memory Mechanisms of Foundation Agents in the Second Half notes that hundreds of memory papers appeared in 2025 because agents face context explosion beyond fixed windows.
Without robust memory substrates, foundation agents in deep research, agentic coding, and computer use cannot reuse long-horizon interaction history, leaving real-world utility far below benchmark scores.
HOW IT WORKS
Rethinking Memory Mechanisms of Foundation Agents in the Second Half introduces a taxonomy over Memory Substrates, Memory Cognitive Mechanisms, and Memory Subjects, plus operation architectures and optimization policies.
You can think of internal memory as RAM, external memory as disk, and episodic or semantic stores as a card catalog that indexes long-term experiences.
This organization lets Rethinking Memory Mechanisms of Foundation Agents in the Second Half explain how memory management itself becomes a trainable capability, something a plain context window cannot provide for long-horizon, user-dependent tasks.
DIAGRAM
This diagram shows how Rethinking Memory Mechanisms of Foundation Agents in the Second Half structures memory operations for single agent and multi agent architectures.
DIAGRAM
This diagram shows how Rethinking Memory Mechanisms of Foundation Agents in the Second Half maps 218 papers across substrates, mechanisms, and subjects over time.
PROCESS
Taxonomy of Memory in Foundation Agents
Rethinking Memory Mechanisms of Foundation Agents in the Second Half first defines Memory Substrates, Memory Cognitive Mechanisms, and Memory Subjects to categorize 218 works systematically.
Memory Operation Mechanism
Rethinking Memory Mechanisms of Foundation Agents in the Second Half then analyzes Memory Operations for Single Agent Systems and Memory Operations for Multi Agent Systems.
Memory Evolution and Optimization Policies
Rethinking Memory Mechanisms of Foundation Agents in the Second Half studies Prompt Driven Memory Evolution and Optimization, Fine Tuning for Parameterized Memory Policies, and Reinforcement Learning for Memory Policies.
Evolving and Scaling Memory Contexts and Environments
Rethinking Memory Mechanisms of Foundation Agents in the Second Half finally connects memory design to Context Limited Simple Environments and Context Exploded Real World Environments plus evaluation and future directions.
KEY CONTRIBUTIONS
Three Dimensional Taxonomy of Memory
Rethinking Memory Mechanisms of Foundation Agents in the Second Half introduces a unified taxonomy over Memory Substrates, Memory Cognitive Mechanisms, and Memory Subjects, connecting internal or external stores with cognitive roles and user or agent focus.
System Level Analysis of Memory Operations
Rethinking Memory Mechanisms of Foundation Agents in the Second Half analyzes Memory Operations for Single Agent Systems and Multi Agent Systems, detailing how memory is routed, updated, and coordinated across agents.
Positioning Memory as Self Evolution Substrate
Rethinking Memory Mechanisms of Foundation Agents in the Second Half highlights learning policies where reinforcement learned context curation and experience consolidation turn memory management into a self evolving capability.
RESULTS
Collected Papers
218 papers
covers 2023 Q1 to 2025 Q4 across memory substrates, mechanisms, and subjects
Time Span
2023 2025
shows rapid acceleration of memory research in 2025
Substrate Categories
2 types
internal and external memory substrates explicitly partitioned
Cognitive Mechanisms
5 types
sensory, working, episodic, semantic, procedural mapped to agent designs
Rethinking Memory Mechanisms of Foundation Agents in the Second Half builds a dataset of 218 memory related agent papers from 2023 Q1 to 2025 Q4, mapped onto substrates, cognitive mechanisms, and subjects. This synthesis shows that memory research surged in 2025 and that external memory and working or episodic mechanisms dominate long horizon, context exploded settings.
BENCHMARK
Rethinking Memory Mechanisms of Foundation Agents in the Second Half builds a dataset of 218 memory related agent papers from 2023 Q1 to 2025 Q4, mapped onto substrates, cognitive mechanisms, and subjects. This synthesis shows that memory research surged in 2025 and that external memory and working or episodic mechanisms dominate long horizon, context exploded settings.
BENCHMARK
Relative distribution of 218 collected papers across memory substrates, cognitive mechanisms, and memory subjects as described in Figure 3.
KEY INSIGHT
Rethinking Memory Mechanisms of Foundation Agents in the Second Half shows that memory management itself is becoming a trainable capability, not just a static engineering choice.
This is surprising because many assumed larger context windows alone would solve long horizon issues, but the survey argues selective storage, reuse, and forgetting policies are essential.
WHY IT MATTERS
Rethinking Memory Mechanisms of Foundation Agents in the Second Half gives builders a vocabulary and design space to mix internal, external, episodic, semantic, and procedural memory for specific deployments.
With this structure, practitioners can design self evolving agents that accumulate experience, share skills, and maintain user centric personalization across long horizon, multi session environments.
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…