memory retention
Deciding what to keep as information arrives, before any query shows up, is memory retention. With bounded context, raw records preserve precise details but limited coverage; consolidation covers more but can lose critical specifics. It also covers objects leaving view, influences persisting after input removal, and stored formats surviving model upgrades. Retention is separate from retrieval: storing versus surfacing later.
Across these studies, no single retention strategy wins. The best choice shifts with budget pressure; compressed summaries can lose accuracy, while raw history often enables recovery. Models perceive dynamic scenes but quickly lose unobserved objects. Attacks persist in the retained cache after input removal. Evaluations mistake extra access for better memory, so retention and selection must be reported separately.
- Papers
- 32
- Released code
- 0
- First seen
- July 2026
- Latest
- Nov 2026
29 papers in the last two months, against 3 in the two before.
Start here
Three papers to read first, in the order they came out- 1Memory Tests Must Separate What Agents Keep From What They Retrieve
Memory evaluations confound retention with retrieval; most apparent query-awareness gains are actually history access, so benchmarks should hold access fixed and report both separately.
Oct 2026 - 2Agent memory can covertly carry secrets across separate sessions
Agentic memory can function as a persistent cross-session covert channel for steganographic attacks.
Oct 2026 - 3Memory-augmented Mamba branch improves GraphCast forecast error by 14 percent
Adding a Mamba memory correction to frozen GraphCast reduces ten-day forecast error by 14.42%.
Oct 2026
How fast it is moving
Often studied with
agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.15 togetherbenchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.5 togethercounterfactual reasoningComparing actual outcomes with imagined alternatives to estimate the effects of different actions or conditions.4 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Independentcs.CR
Agent memory can covertly carry secrets across separate sessions
Agentic memory can function as a persistent cross-session covert channel for steganographic attacks.
Oct 2026
- Top Universitycs.AI
Memory Tests Must Separate What Agents Keep From What They Retrieve
Memory evaluations confound retention with retrieval; most apparent query-awareness gains are actually history access, so benchmarks should hold access fixed and report both separately.
Stanford University · Oct 2026
- Big Techcs.LG
Memory-augmented Mamba branch improves GraphCast forecast error by 14 percent
Adding a Mamba memory correction to frozen GraphCast reduces ten-day forecast error by 14.42%.
Princeton University, École Polytechnique Fédérale de Lausanne (EPFL) · Oct 2026
- Big Techcs.CV
Persistent memory is essential for coherent autoregressive video generation
Memory in autoregressive video generation is a five-part problem: forms, functions, operations, learning, and evaluation.
HKUST, CityUHK · Oct 2026
- Chinese Techcs.AI
Self-distillation choices govern learning gains and preservation of prior abilities
Self-distillation's three axes each affect acquisition and retention differently depending on task contradiction and model identity.
Huawei Technologies Co., Ltd. · Oct 2026
- Independentcs.CR
Adversarial images and prompts steer models after being removed from context
Adversarial inputs can persist through the KV cache, continuing to influence a model after being removed from the context.
Oct 2026
- Academiccs.CV
4D foundation models lack robust visual memory for out-of-view objects
Current 4D foundation models fail to remember objects that exit their field of view.
Cornell University · Sept 2026
- Independentcs.LG
AI agents track permissions better from updated records than long histories
Maintaining an explicit state for an agent beats making a language model read the full event log, past a log length that grows with the model's strength.
Nov 2026
- Academiccs.AI
Randomized retrieval reveals which AI memories actually help
CMP uses randomized retrieval to make memory utility identifiable in memory-augmented LLMs.
Illinois Institute of Technology · Oct 2026
- Big Techcs.AI
LAM Cuts Agent Memory While Bounding Retrieval Score Errors
LAM trims 22.47% of agent memory tokens with 99.984% evidence retention and predicts up to 91.6x speedup by scheduling compaction before prefill.
Argonne National Laboratory, University of Houston · Sept 2026
- Independentcs.AI
Structured agent memories withstand model upgrades better than compressed notes
Fixed-schema knowledge graphs transfer across model upgrades without accuracy loss, while compressed natural-language notes lose up to 13 percentage points.
Sept 2026
- Independentcs.AI
Memory consolidation outperforms retention under tight budgets, but retention wins when budgets are loose.
When evidence is long relative to the token budget, consolidating memory with Merge or Abstract improves accuracy; under loose budgets, retaining raw records is better.
July 2026
- Chinese Techcs.LG
AI agents can recover forgotten skills from earlier training snapshots
Agentic-OPD recovers lost capabilities from earlier RL checkpoints via on-policy distillation, improving agent performance beyond the best checkpoint.
Baidu, Inc., Beijing Jiaotong University · Nov 2026
- Chinese Techcs.CV
AI agents can perform better while keeping fewer memories
Memory curation by replaying source decisions before retention improves LLM agent success rates and reduces token use.
The University of Hong Kong, Huazhong University of Science and Technology · Nov 2026
- Big Techcs.CV
AI systems reconstruct scenes with less drift using recent history
HIST3R controls the direction and magnitude of state updates using history to reduce drift in long-video 3D reconstruction.
National Taiwan University, NVIDIA · Nov 2026
- Independentcs.AI
AI agents finish tasks normally while quietly following planted instructions
Malicious instructions retained in an AI agent’s memory can quietly redirect later decisions while the agent still completes legitimate tasks normally.
Nov 2026
- Big Techcs.CV
One vision backbone can continually absorb new foundation model capabilities
GRAFT continually extends a vision backbone by distilling from new teachers while using the previous model to retain old capabilities.
University of Illinois Chicago, Amazon · Oct 2026
- Independentcs.LG
Adaptive memory graphs improve collaboration in multi-agent language-model systems
MACE adapts memory organization and agent memory use through feedback, outperforming ten baselines on eight benchmarks.
Sept 2026
- Big Techcs.LG
Language models' accuracy varies with update timing, not just update counts
The Control Map evaluates test-time updates by recording update histories and separating acquisition, retention, and locality.
Meta AI · Oct 2026
- Chinese Techcs.AI
AI agents fail tasks despite recording correct knowledge about environments
Agents that keep notes from past tasks do better, but they fail to update those notes when the environment quietly changes, and sometimes still fail even when the notes are right.
The Hong Kong University of Science and Technology, Huawei Inc. · Nov 2026
- Industrycs.CR
Persistent memory in LLMs enables cross-session security breaches
Memory injection attacks persistently compromise ChatGPT Instant 5.3, Claude Sonnet 4.6, and Gemini 3 Flash, achieving all NIST attack goals through stored prompts.
Universidade de São Paulo (USP), Universidade do Estado de Santa Catarina (UDESC) · Oct 2026
- Independentcs.AI
Discarding stored abstractions improves how LLM agents learn from experience
Persisting inferred agent memories can reproduce errors, while rebuilding temporary guidance from stored trajectories improves held-out success and reduces LLM-call cost.
Oct 2026
- Big Techcs.LG
Teacher training determines whether on-policy distillation preserves prior language skills
RLVR-trained distillation teachers preserve prior capabilities better than SFT-trained teachers, while out-of-domain predictive change tracks forgetting most consistently.
Microsoft Research, Nankai University · Oct 2026
- Big Techcs.AI
New benchmark tests MLLM agents on dense, stateful professional visuals
The best MLLM agent scores below 45% on the DSV-Mem benchmark for professional multimodal memory.
Amazon AGI, University of Southern California · Oct 2026
- Chinese Techcs.CV
Benchmark exposes separate static and dynamic memory failures in video worlds
WorldMemBench evaluates video world models on static and dynamic memory retention across repeated revisits, finding no model leads all dimensions.
Zhejiang University, Ant Group · Oct 2026
- Academiccs.CV
Video world model uses test-time training as persistent memory for long-range consistency
A camera-controlled video world model uses test-time training memory to preserve observed scenes in fixed-size parameters, improving long-horizon revisit consistency.
Meta Reality Labs, Brown University · Oct 2026
- Research Labcs.AR
Unified retention and compensation for KV cache eviction via coverage calibration
CORE unifies KV-cache retention and compensation by calibrating one allocation distribution that jointly selects Top-B states and writes latent memory from excluded mass.
Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences · Oct 2026
- Top Universitycs.CL
Counterfactual credit teaches language models which memories stay useful
MGPO isolates incremental value of memory rewrites via counterfactual credit, reducing memory length by 80% while improving extraction.
University of Michigan · Sept 2026
- Top Universitycs.AI
Coding Agents Need Semantic Measures of Working Memory Performance
Semantic heterogeneity in coding-agent working memory requires evaluation beyond nominal token budgets.
Argonne National Laboratory, Columbia University · Sept 2026
- Big Techcs.CL
Voice agents lose document grounding as conversations and context grow
Document-grounded voice agents lose factual fidelity as context and conversation grow, often hallucinating instead of abstaining; cascaded systems ground best among those evaluated.
Adobe Research, University of Maryland College Park · Oct 2026
- Big Techcs.RO
Recurrent memory policy improves robot manipulation on tasks requiring spatial recall, counting, and timing
ReCAT uses structured recurrent memory with Mamba-2 and causal attention to achieve high success on memory-dependent robot manipulation tasks.
University of Edinburgh, Karlsruhe Institute of Technology · Oct 2026
- Top University
Structured memory helps health agents interpret longitudinal patient conversations
PIA turns health conversations into structured clinical records and synthesizes user understanding with four memory controls.
KAIST, NAVER Corp. · Sept 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.