# Agent Memory

30 related articles

Subscribe to updates

arXiv Nov 4, 2026
Reglos splits personal-agent memory into two loops: memory grows with each chat while a coding agent rewrites the harness
Reglos, a personal memory agent, splits improvement across two time scales: each user's memory grows online as conversations arrive, while the harness that decides which memories reach the model is revised offline by a coding agent. A critic from a second model family screens patches for benchmark gaming, and a fixed evaluator accepts them only on users held out from development. With Qwen3.8-27B writing memory and answering, Reglos reaches 61.3% on PersonaMem-v2, 7.5 points above the strongest compared system.
Reglos, a personal memory agent, splits improvement across two time scales: each user's memory grows online as conversations arrive, while the harness that decides which memories reach the model is revised offline by a coding agent. A critic from a second model family screens patches for benchmark gaming, and a fixed evaluator accepts them only on users held out from development. With Qwen3.8-27B writing memory and answering, Reglos reaches 61.3% on PersonaMem-v2, 7.5 points above the strongest compared system.
Sen Yang（First author）
Agent Memory Recursive Self-Improvement Reglos Self-Evolving Agents PersonaMem-v2
arXiv Nov 4, 2026
Agents write their own rules into handoff summaries: one clearer approval sentence in the system prompt cuts unsupported restrictions from 5.0% to 1.8%
Long-running LLM agents compact their context by writing a handoff summary that the next instance continues from, and that summary can state permissions or requirements the record does not support. Using synthetic tool-use tasks with a ground-truth permission ledger and same-snapshot replay, two AI readers confirmed unsupported restrictions in 1.2% and 2.4% of readable summaries from two open-weight models on held-out instances, and unsupported authorizations in 0.1% and 1.3%. Replacing the system prompt's "Some actions need approval." with a sentence saying where requirements come from cut unsupported restrictions from 5.0% to 1.8% of development summaries and from 1.8% to 0.3% on held-out templates, while requests for approvals the ledger requires barely changed.
Long-running LLM agents compact their context by writing a handoff summary that the next instance continues from, and that summary can state permissions or requirements the record does not support. Using synthetic tool-use tasks with a ground-truth permission ledger and same-snapshot replay, two AI readers confirmed unsupported restrictions in 1.2% and 2.4% of readable summaries from two open-weight models on held-out instances, and unsupported authorizations in 0.1% and 1.3%. Replacing the system prompt's "Some actions need approval." with a sentence saying where requirements come from cut unsupported restrictions from 5.0% to 1.8% of development summaries and from 1.8% to 0.3% on held-out templates, while requests for approvals the ledger requires barely changed.
University at Albany, State University of New York State University of New York
W. J. Chu（First author, Corresponding author）
Handoff summaries Context Compaction Same-snapshot replay LLM Agents Agent Memory
arXiv Nov 2, 2026
Personal memory updates split in two: MemShift scores missed changes apart from wrong repairs
Values an AI assistant derives from a person's records, such as important contacts, can go stale when an input changes. MemShift uses synthetic worlds that fork only on one kind of change to score whether a memory noticed the change and whether its repair was right, against gold fixed before any text and without an LLM judge. Across four open-weight models, detection helps only as much as repair allows: after a redefinition, the key-by-exact-recomputation interaction is +38.8 percentage points.
Values an AI assistant derives from a person's records, such as important contacts, can go stale when an input changes. MemShift uses synthetic worlds that fork only on one kind of change to score whether a memory noticed the change and whether its repair was right, against gold fixed before any text and without an LLM judge. Across four open-weight models, detection helps only as much as repair allows: after a redefinition, the key-by-exact-recomputation interaction is +38.8 percentage points.
The University of Osaka Nagoya University
Zichuan Xu（First author）
MemShift Agent Memory Detection and Repair of Derived Memory HoldMem
arXiv Nov 4, 2026
Agent memory gets its own compressor: SumOne learns from masked reconstruction and cuts search-task cost to 22% of full context
Long-horizon language-model agents accumulate ever-longer interaction histories, so context compaction is necessary, yet existing learned compaction methods often depend on downstream task outcomes or solver behavior. SumOne is a standalone 2-billion-parameter summarizer trained with a task-agnostic reward built from masked reconstruction and optimized with GRPO, using only raw trajectories and no reference summaries or downstream task execution. On DeepSearchQA it cuts cumulative agent input from 2.83M to 591K tokens with F1 0.48 versus 0.52 for full context; on SWE-bench Verified it reaches a 0.59 resolved rate versus 0.63, with input falling from 4.10M to 1.09M tokens.
Long-horizon language-model agents accumulate ever-longer interaction histories, so context compaction is necessary, yet existing learned compaction methods often depend on downstream task outcomes or solver behavior. SumOne is a standalone 2-billion-parameter summarizer trained with a task-agnostic reward built from masked reconstruction and optimized with GRPO, using only raw trajectories and no reference summaries or downstream task execution. On DeepSearchQA it cuts cumulative agent input from 2.83M to 591K tokens with F1 0.48 versus 0.52 for full context; on SWE-bench Verified it reaches a 0.59 resolved rate versus 0.63, with input falling from 4.10M to 1.09M tokens.
Peking University
Jizhe Zhang（First author）
SumOne Masked Reconstruction Reward GRPO Agent Memory Context Compaction
arXiv Nov 10, 2026
Long-horizon agents keep memory in plain files: CAMP-II averages 64.6% across five environments, above the 54.2% task-reward reference CAMP-I
Long-horizon agents accumulate interaction histories that outgrow the active context, so they must learn what to write down and when to read it back. Earlier on-policy distillation of memory control relied on dedicated memory tools or compaction protocols that neither the teacher nor the student had seen, which weakened guidance on the teacher side and left the student sampling too few meaningful memory actions. CAMP-II instead routes memory through ordinary Bash file operations that both models already have a prior for, training one policy jointly across five long-horizon environments to a 64.6% average task success rate against 54.2% for the task-reward reference CAMP-I, with transfer to Terminal-Bench 2.1 and SWE-Bench Pro V2.
Long-horizon agents accumulate interaction histories that outgrow the active context, so they must learn what to write down and when to read it back. Earlier on-policy distillation of memory control relied on dedicated memory tools or compaction protocols that neither the teacher nor the student had seen, which weakened guidance on the teacher side and left the student sampling too few meaningful memory actions. CAMP-II instead routes memory through ordinary Bash file operations that both models already have a prior for, training one policy jointly across five long-horizon environments to a 64.6% average task success rate against 54.2% for the task-reward reference CAMP-I, with transfer to Terminal-Bench 2.1 and SWE-Bench Pro V2.
Lirui Luo（First author）
CAMP-II CAMG-II On-Policy Distillation Agent Memory LLM Agents
MongoDB | Blog Nov 4, 2026
MongoDB and Google ADK connect: agent sessions and memory no longer need a separate store
MongoDB released an integration with Google ADK that gives agent state durable memory. Previously developers still had to build their own tooling for a data store that could retain agent context over time. The integration uses MongoDbSessionService to persist sessions, events, and application-, user-, and session-scoped state to MongoDB, and MongoDBMemoryService to ingest session events into collections and recall memories through keyword matching; retrieval reuses MongoDB vector search combined with full-text search. It supports MongoDB Atlas and self-managed MongoDB 8.0+ deployments.
MongoDB released an integration with Google ADK that gives agent state durable memory. Previously developers still had to build their own tooling for a data store that could retain agent context over time. The integration uses MongoDbSessionService to persist sessions, events, and application-, user-, and session-scoped state to MongoDB, and MongoDBMemoryService to ingest session events into collections and recall memories through keyword matching; retrieval reuses MongoDB vector search combined with full-text search. It supports MongoDB Atlas and self-managed MongoDB 8.0+ deployments.
MongoDB
MongoDB Google ADK Integration Agent Memory MongoDB Vector Search
arXiv Oct 14, 2026
Streaming video agents stop recording only what is present: DeltaStreamer logs frame-to-frame changes and scores 66.07% on OVO-Bench
Streaming video understanding must retain evidence before anyone knows which details later queries will need, yet existing methods flatten event transitions when compressing history and cannot recover details once a frame has passed. DeltaStreamer explicitly records visual changes between adjacent frames in a hierarchical textual memory, and after a query arrives a query-driven planner decides whether to retrieve, keep observing, or replay historical frames. It scores 66.07% overall on OVO-Bench and 59.53% on StreamingBench, above the representative streaming methods it is compared with.
Streaming video understanding must retain evidence before anyone knows which details later queries will need, yet existing methods flatten event transitions when compressing history and cannot recover details once a frame has passed. DeltaStreamer explicitly records visual changes between adjacent frames in a hierarchical textual memory, and after a query arrives a query-driven planner decides whether to retrieve, keep observing, or replay historical frames. It scores 66.07% overall on OVO-Bench and 59.53% on StreamingBench, above the representative streaming methods it is compared with.
Nankai University Chinese Academy of Sciences ByteDance (China)
Boyuan Sun（First author）
DeltaStreamer Streaming Video Understanding Agent Memory Query-Driven Planning
arXiv Oct 20, 2026
Long-term memory becomes callable code: MaC lifts overall LLM-judge accuracy by up to 15.91% on LOCOMO while using fewer tokens
Long-term memory systems usually store facts as entries and pick them by relevance at query time, which makes facts scattered across sessions and changing over time hard to compose. MaC puts memory construction and retrieval under one code abstraction: facts become executable functions with canonical identifiers and source references, and a query is compiled into a retrieval program that an executor runs after a single planning call. On LOCOMO and LONGMEMEVAL, MaC reaches the highest overall LLM-judge score in all four evaluation settings, improving over the strongest baseline in each by up to 15.91% and 9.06%, while substantially cutting token consumption.
Long-term memory systems usually store facts as entries and pick them by relevance at query time, which makes facts scattered across sessions and changing over time hard to compose. MaC puts memory construction and retrieval under one code abstraction: facts become executable functions with canonical identifiers and source references, and a query is compiled into a retrieval program that an executor runs after a single planning call. On LOCOMO and LONGMEMEVAL, MaC reaches the highest overall LLM-judge score in all four evaluation settings, improving over the strongest baseline in each by up to 15.91% and 9.06%, while substantially cutting token consumption.
Shanghai Jiao Tong University Chinese University of Hong Kong, Shenzhen UNSW Sydney
Aoting Zeng（First author）
Memory as Code (MaC)Declarative Retrieval Codified Construction Agent Memory LoCoMo
researchsquare.com Nov 5, 2026
A personal-health agent that remembers you: HealthJev screens bad lessons in 86 ms and lifts six tasks by 13.3 points on average
Medical large language models answer single encounters well but struggle to accumulate and govern experience across encounters. HealthJev separates shared safety rules and medical knowledge from a private longitudinal memory, and a self-trained lightweight decision model, JevGate, screens candidate lessons before they enter memory at a median 86 ms per decision. On a longitudinal benchmark of 90 simulated daily interactions for 20 users built from real MIMIC-IV trajectories, it scored a mean 0.668, versus 0.634 for MemoryBank, 0.593 for Mem0, 0.510 for ReAct and 0.502 for Memp; across six biomedical tasks it gained 13.3 percentage points on average over the base condition.
Medical large language models answer single encounters well but struggle to accumulate and govern experience across encounters. HealthJev separates shared safety rules and medical knowledge from a private longitudinal memory, and a self-trained lightweight decision model, JevGate, screens candidate lessons before they enter memory at a median 86 ms per decision. On a longitudinal benchmark of 90 simulated daily interactions for 20 users built from real MIMIC-IV trajectories, it scored a mean 0.668, versus 0.634 for MemoryBank, 0.593 for Mem0, 0.510 for ReAct and 0.502 for Memp; across six biomedical tasks it gained 13.3 percentage points on average over the base condition.
Huazhong University of Science and Technology Fudan University Jingdong (China)Nanjing Medical University Beijing University of Chinese Medicine
HealthJev JevGate Agent Memory Longitudinal Personal Health Benchmark LLM Agents
arXiv Nov 1, 2026
Delivered memory is not used memory: MemUse traces each delivery and finds most entries irrelevant and most relevant ones ignored
MemUse moves the unit of evaluation from the task outcome to the individual memory delivery: across 8,000 SWE-bench Verified agent episodes, three memory systems stayed within 0.4 percentage points of a no-memory baseline on resolve rate, yet per-delivery judgments show most delivered entries were irrelevant, 51.3%–79.7% of relevant entries went unused in each system, and 23.2% of episodes with beneficial memory use still failed.
MemUse moves the unit of evaluation from the task outcome to the individual memory delivery: across 8,000 SWE-bench Verified agent episodes, three memory systems stayed within 0.4 percentage points of a no-memory baseline on resolve rate, yet per-delivery judgments show most delivered entries were irrelevant, 51.3%–79.7% of relevant entries went unused in each system, and 23.2% of episodes with beneficial memory use still failed.
Changran Xu（First author）
MemUse Agent Memory Agent evaluation SWE-bench Verified LLM-as-a-Judge