Giving an AI better memory of what you like doesn't mean it will actually act on it.
Do memory architectures genuinely close the gap between knowing and acting on preferences?
This explores whether giving agents better memory (storing, retrieving, and compressing what a user likes) makes them behave according to those preferences, or only makes them remember the preferences more accurately.
This explores whether better agent memory makes agents act on what they know about you, or just makes them better at recalling it. The corpus's most direct evidence says the gap survives. Paired "Know" and "Act" tests across 16 systems found that agents pass recall tests but fail to show the preference in their behavior, and the failures come mostly from misinterpreting the memory, not from failing to retrieve it Why do LLM agents remember preferences but not act on them?. An agent can pull up "I'm vegetarian" perfectly and still book the steakhouse.
That is a problem for much of the memory literature, because most of it works on storage and retrieval. Titans stores surprising tokens in a separate neural memory and scales past two million tokens of context Can neural memory modules scale language models beyond attention limits?. MeMo trains a dedicated memory model so new knowledge gets in without touching the main LLM or searching a corpus at inference time Can a separate memory model inject knowledge without touching the LLM?. Memory folding compresses interaction history into structured schemas Can agents compress their own memory without losing critical details?. These give an agent cheaper and larger memory, but they are measured on capacity, cost, and task success. None of their summaries tests whether the model applies a remembered preference. They improve the stage that the Know/Act result says is not the bottleneck.
A few designs plausibly reach the applying step because they change the form of memory, not its size. PRIME finds that abstract preference summaries beat retrieved specific past interactions, and that recency beats similarity for recall Does abstract preference knowledge outperform specific interaction recall?. A summary has already been interpreted, so the model has less inference to do when it decides. MRAgent interleaves reasoning with memory traversal instead of retrieving first and reasoning afterward, and gains up to 23% on reasoning tasks Can agents reconstruct memory on demand instead of retrieving it?. SkillRL keeps successes as concrete demonstrations and turns failures into abstracted lessons Should successful and failed episodes be processed differently?. The common thread is handing the model memory that has been pre-digested into something it can act on. This is a hypothesis that fits the Know/Act finding. These studies report task performance, not a before-and-after comparison of recall against behavior.
Some systems do act on what they remember. M3-Agent builds an entity-centric graph that separates episodic events from semantic knowledge, and runs memorization and control in parallel so it can infer and act on preferences from observation without asking Can agents learn preferences by watching rather than asking?. AgentFly improves entirely through memory operations, with no weight updates, and reaches 87.88% on GAIA validation Can agents learn continuously from experience without updating weights?. Both show that memory can drive behavior. Neither is described as testing the recall-versus-action gap.
The corpus's evaluation work explains why the gap is easy to miss. Scoring memory only end-to-end hides which component failed. Breaking it into storage, extraction, retrieval, and maintenance exposes trade-offs that task scores conceal How should we actually evaluate agent memory systems?. Those four stages have no slot for applying the memory, which is where the Know/Act failures occur. So the answer so far is no, not demonstrably. The gap is well documented, the fixes are mostly indirect, and the field's own measurements are not yet built to show whether it has closed.
Sources 10 notes
Paired Know and Act tests across 16 systems revealed a large gap: agents pass recall tests but fail to reflect preferences in behavior. Comprehension failures during interpretation dominate over retrieval failures, suggesting the bottleneck lies in applying stored information rather than retrieving it.
Titans architecture separates attention (short-term, quadratic) from neural memory (long-term, compressed), prioritizing surprising tokens for storage. The model outperforms standard Transformers and linear RNNs across tasks while scaling to 2M+ token contexts without quadratic penalties.
MeMo trains a dedicated memory model to encode new knowledge, eliminating inference-time search costs that scale with corpus size. It avoids fine-tuning risks and works with frozen proprietary models, but trades this for up-front training cost and capacity limits.
DeepAgent's autonomous memory folding consolidates interaction history into episodic, working, and tool memory schemas. This reduces token overhead while letting agents pause to reconsider strategies—the autonomy and structure together avoid degradation that plagues poorly designed consolidation.
PRIME framework shows semantic memory (preference summaries, parametric encodings) consistently beats episodic memory (retrieved past interactions) across models. Recency-based recall outperforms similarity-based retrieval, and task fine-tuning exceeds preference tuning methods.
Show all 10 sources
MRAgent achieves up to 23% gains on reasoning tasks by reconstructing memory through active graph traversal that prunes paths based on accumulated evidence, while reducing token and runtime cost compared to fixed-retrieval pipelines.
SkillRL demonstrates that treating successful episodes as concrete demonstrations and failures as abstracted lessons achieves state-of-the-art performance on complex tasks while using substantially less context than uniform approaches. The asymmetry mirrors human expert reasoning and avoids the degradation seen in uniform consolidation methods.
M3-Agent demonstrates that separating episodic events from semantic knowledge in an entity-centric graph, combined with parallel memorization and control processes, allows agents to infer and act on user preferences without asking. This architecture mirrors human cognitive systems that bind disparate information about individuals across sensory modalities.
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
Decomposing memory into storage, extraction, retrieval, and maintenance stages exposes design trade-offs and failure modes that task-success metrics completely hide. Module-by-module evaluation across 12 systems shows which component actually failed rather than just whether the task succeeded.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- Useful Memories Become Faulty When Continuously Updated by LLMs
- Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs
- Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents
- SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
- Rethinking Memory as Continuously Evolving Connectivity