Research · NVIDIA Technical Blog
NemoClaw: memory improved accuracy, but sometimes got in the way
In NVIDIA’s experiment, structured memory improved a team agent’s answers, especially on difficult questions. Some metrics nevertheless declined. The result clarifies what deserves investment when preparing a local assistant: connections between decisions, traceable sources, and user corrections that survive subsequent runs.
What this means for a buyer
For a team assistant, organizing memory may deserve funding before another increase in model size. Before choosing a more powerful server, identify which decisions must persist, how current instructions differ from old ones, and who corrects the accumulated view of the work.
A conversation preserves words; a project needs connections
In an NVIDIA article dated September 4, 2026, the authors compared an agent that retrieves source material with one using structured memory. Accuracy across 186 questions rose from 82.8% to 90.9%, an increase of 8.1 percentage points. Both configurations used NVIDIA Nemotron 3 Ultra. The interesting development concerns the organization of working context: preserved relationships between events improved answers, although some individual categories became worse.
Conversation history mixes current priorities, withdrawn requests, and intermediate decisions. Retrieval can find the right messages, but those messages still need to be connected. In a hypothetical project, for example, a deadline is agreed, then postponed, while the project name is later shortened. An answer about its current status depends on all three relationships. An assistant that finds only the original deadline can cite a real document and still be wrong.
Memory that can be inspected
In the NemoClaw example, the authors call this memory the self model. It consists of structured Markdown pages about people, projects, priorities, and working patterns. Its schema includes indexing, cross-references, provenance, and growth limits. For each task, the agent receives a bounded selection of relevant context. This makes it possible to discuss a particular page or relationship when the assistant misunderstands a project’s status.
Those pages contain a derived interpretation of the work. Original messages retain their role as evidence. A separate SQLite ledger stores obligations, importance judgments, corrections, and audit events. The agent does not write its judgments into the sources as read flags or labels. That separation makes it possible to investigate an error in stages: incorrect evidence, a faulty memory update, poor retrieval, or a mistaken final decision.
The improvement was uneven
The Agent Memory Benchmark comparison used an agentic RAG baseline, which retrieves material over multiple rounds, and a configuration with the self model. Accuracy on 31 hard questions rose from 67.7% to 87.1%. This is particularly interesting for teamwork, where an answer often depends on several messages. The overall result measures answer quality on this question set, rather than server speed or the number of employees it can serve simultaneously.
There was a downside. In the category measuring answers faithful to the corpus, containing 13 questions, the result fell from 100.0% to 92.3%. On 30 single-hop lookup questions, accuracy declined from 86.7% to 83.3%. Small categories deserve cautious interpretation, but the regressions still matter. For a document assistant, they support preserving access to original evidence: even a well-organized interpretation can sometimes perform worse than directly finding a particular fact.
A correction becomes part of the next decision
Persistent memory can preserve a mistaken judgment for a long time. In the recipe, users can move an obligation to another priority tier or ignore it; subsequent runs retain that decision. Each change is recorded once in an append-only audit trail. Repeated corrections can update a small, readable preference policy that users can inspect, edit, or delete.
The authors also separate the urgency claimed by an incoming message from the user’s intent. The highest tier is reserved for obligations connected to stated priorities, while ordinary code limits tier sizes and enforces ranking order. This is a meaningful feature for a team: an assistant should remember why a task matters and respect an explicit correction from its owner. Otherwise, accumulating context can simply make an incorrect recommendation more persistent.
Where memory ends and action begins
NemoClaw integrates the example with the NVIDIA OpenShell secure runtime. It restricts the agent’s access to files, processes, and the network. For managed model inference and MCP connections, credentials remain outside the sandbox. A memory entry about a preferred communication channel can inform a suggested action, while authorization to send remains separate. Memory and retrieved documents continue to be inputs to the model.
The current recipe addresses the memory foundation: it neither sends messages nor changes source systems. Its people and messages are invented, and the offline walkthrough uses recorded model decisions in place of live inference. A real team assistant therefore needs a separate preparation budget for organizing and maintaining memory. NVIDIA’s result makes that preparation a justified first investment before trying to resolve these errors solely by replacing the model.
What to check before choosing
- Which project renamings and deadline changes must the assistant connect?
- Can a memory page open the original message and show how its conclusion was derived?
- Who can correct priorities and delete stored preferences?
- Is integration with workplace systems needed now, or would a reading assistant suffice?
Buyer implications are our editorial interpretation. The publication does not establish the speed or compatibility of a particular catalogue system.