NVIDIA NemoClaw Architecture for Building Memory-Driven Enterprise Agents

NVIDIA NemoClaw Architecture for Building Memory-Driven Enterprise Agents NVIDIA outlined a system design using NVIDIA NemoClaw and NVIDIA OpenShell to create a memory-driven Chief of Staff agent that leverages persistent context, separates knowledge from execution, and maintains strict runtime governance. The system separates raw evidence, structured knowledge stored in a Markdown 'self model', and agent judgments tracked in a SQLite ledger. An intent gate prioritizes tasks based on stated user goals rather than incoming urgency, while an append-only audit trail records user corrections to update readable preference policies. Benchmark evaluations using NVIDIA Nemotron 3 Ultra showed the self model increased overall accuracy to 90.9% from an 82.8% agentic RAG baseline, including a 40 percentage point boost in tracking changing facts. Runtime security is enforced via NVIDIA OpenShell sandboxing, which isolates credentials and enforces file, process, and network boundaries independently of agent context. Structuring Memory and Separating Knowledge from Action Enterprise agents require continuous context rather than fleeting conversation history. NemoClaw achieves this through a human-readable 'self model' composed of structured Markdown pages that catalog people, projects, priorities, and recurring patterns without overwriting source evidence. A separate SQLite ledger logs obligations, rankings, corrections, and audit events. This three-layer design—evidence, knowledge, and governed execution—ensures context guides recommendations without directly granting authorization to execute actions like sending messages. Intent-Based Prioritization and User Correction Loops Incoming communications frequently present artificial urgency that diverges from user priorities. NemoClaw employs an intent gate to rank tasks according to explicitly stated user goals, placing quiet high-priority tasks above ostensibly urgent requests like expense attestations. When an agent errs, users can re-tier or dismiss obligations; each intervention enters an append-only audit trail. Repeated adjustments update a transparent, user-editable preference policy rather than embedding hidden biases into model weights. Agent Memory Benchmark Performance Across a 186-question benchmark evaluating NVIDIA Nemotron 3 Ultra, the self model outpaced an agentic multi-round retrieval baseline by raising overall accuracy from 82.8% to 90.9%. It achieved notable improvements in hard questions (87.1% versus 67.7%), point-in-time reasoning (66.7% versus 33.3%), entity disambiguation (86.7% versus 66.7%), and tracking facts that changed over time (100.0% versus 60.0%). Minor regressions occurred in single-hop lookups (83.3% versus 86.7%) and faithful corpus answering (92.3% versus 100.0%). Runtime Boundaries and Implementation Scope Because agent memory constitutes unverified input rather than trusted policy, execution safety relies on NVIDIA OpenShell. OpenShell confines the agent within an isolated sandbox and regulates file, process, and network permissions, keeping Model Context Protocol and inference credentials outside the runtime. An open-source Hermes profile recipe is available in the NVIDIA/nemoclaw-community repository with synthetic data and offline walkthroughs to test ranking and memory logic before integrating live enterprise connectors.

英伟达推出基于 NemoClaw 的记忆驱动型智能体架构 英伟达团队利用 NVIDIA NemoClaw 构建了记忆驱动型“幕僚长”智能体，通过自模型结构维护工作上下文，在基准测试中将任务整体准确率提升至90.9%。 采用结构化Markdown“自模型”管理人员、项目与优先级知识，并以SQLite账本记录职责、排序与审计日志。 基准测试显示，相较于Agentic RAG基线，自模型整体准确率提高8.1个百分点，时间演变事实追踪准确率提升40个百分点。 通过意图门控优先保障用户既定目标而非表面紧急事件，并支持用户通过纠错闭环更新可读的偏好策略。 结合 NVIDIA OpenShell 在沙箱中实施安全策略与权限边界，将凭证隔离在沙箱之外以防止越权执行。 证据、知识与执行的分层解耦 该方案将交互过程划分为证据、知识与受管辖的执行三层。系统通过结构化Markdown自模型记录衍生解释而非直接替代源证据，知识涵盖人员、项目与工作模式，而职责、优先级排序和纠错则保存在SQLite账本中。上下文仅为行动提供参考，发送消息或调用工具依然严格依赖凭证、工具权限与用户批准。 意图优先与透明的纠错机制 针对外部请求往往自称紧急但偏离核心目标的现象，系统利用意图门控确保关联用户既定优先级的事项位列最高层级。当智能体出现错误判断时，用户可直接调整层级或忽略任务，所有变更被记入只追加审计追踪中，重复的修正模式会转化为可读的偏好策略供用户审阅、编辑或删除。 基准测试表现与运行时安全边界 在以 NVIDIA Nemotron 3 Ultra 进行的评测中，自模型在硬问题（87.1%）、时点推理（66.7%）及实体消歧（86.7%）等多项任务上显著超越多轮检索基线。运行时则由 NVIDIA OpenShell 提供沙箱环境，隔离文件系统、进程与网络访问，防止模型因误读记忆或遵从恶意指令而突破操作员设定的安全防线。

Enterprise work spans messages, decisions, projects, and obligations that change over time. An AI agent that starts without this context must reconstruct it before contributing. To provide agents with this necessary context, our team used NVIDIA NemoClaw to build a memory-driven Chief of Staff . It maintains a human-readable knowledge layer called the self model : an agent memory of relevant people, projects, priorities, and working patterns. Scheduled jobs periodically review new activity, track obligations, and incorporate user decisions over time. Our experience shows that useful agent memory requires structure, selective retrieval, and governance—not just storage. This post demonstrates how a memory-driven agent built with NVIDIA NemoClaw can improve productivity in real-world enterprise workflows. It also shares five design lessons you can apply to your own agents: Maintain context across daily work to improve task quality Separate evidence, knowledge, and actions to help agents make better judgments Design memory-driven agents can prioritize user intent over short-term urgency Allow users to correct agent decisions to build trust Enforce the boundary of security and authorization with NVIDIA OpenShell Maintain context across daily work Conversation history provides short-term continuity, but it mixes current priorities with past decisions and temporary requests. Retrieval can find relevant source material, but the agent must still connect information across time. Consider a project status question. The answer might depend on an earlier decision, a correction in a later message, an unresolved obligation, and the knowledge that two different names refer to the same project. To solve these problems, you can use the self model that maintains these relationships in structured Markdown pages. It organizes information about people, projects, priorities, goals, concepts, and recurring work patterns. Its schema defines indexing, cross-references, provenance, and growth limits. The self model stores a derived interpretation rather than replacing source evidence. Keeping the two separate helps you, the developer, determine whether an incorrect answer came from the evidence, memory maintenance, retrieval, or the model’s final decision. Separate evidence, knowledge, and action You could use the three layers of tuning capabilities from the memory-driven Chief of Staff: Evidence → Knowledge → Governed execution Evidence supports updates to the self model. For each task, the agent retrieves a bounded set of relevant context. The agent then uses that context within the NVIDIA NemoClaw example. The architecture is shown in Figure 1. Figure 1. NVIDIA NemoClaw uses persistent context for governed agent actions The example stores two kinds of information: Knowledge : People, projects, priorities, and working patterns Judgment : Whether an item needs attention, where it ranks, and whether the user ignored it A Markdown agent memory stores knowledge. A SQLite ledger stores obligations, rankings, corrections, and audit events. This design preserves the agent’s judgments without writing them into source messages as read flags, labels, or folders. A memory page might say that a collaborator prefers Slack. The agent can use that context to recommend Slack, but sending a message still depends on credentials, tool permissions, runtime policy, and user approval. Context can inform an action, but it cannot authorize one. Prioritize user intent over short-term urgency Incoming requests often describe themselves as urgent, but urgency does not necessarily reflect the user’s priorities. An intent gate therefore reserves the highest tier for obligations connected to the user’s stated priorities. In the public recipe provided, an urgent expense-policy attestation remains visible but ranks below a quieter request tied to a stated priority. Your NVIDIA NemoClaw could interpret that relationship, while deterministic code enforces tier size, overflow behavior, and ranking order. Allow users to correct the agent Persistent memory can preserve an incorrect judgment as easily as a correct one. Using the recipe, you could move an obligation to another tier or ignore it, and later agent runs preserve that decision. Each change is recorded once in an append-only audit trail. Repeated correction patterns can update a small, readable preference policy. Users can inspect, edit, or delete that policy instead of leaving the preference hidden in model state. The feedback loop remains visible: Agent judgment → User correction → Audit event → Preference update Add memory to improve agent task performance Adding the memory-driven Chief of Staff to NemoClaw produced measurable improvements across several agent tasks, as shown in Table 1. The Agent Memory Benchmark and evaluation examples are included in the example repo. The example repo compares an agentic retrieval-augmented generation (RAG) baseline that performs multi-round retrieval with the self model. Metrics Question counts Agentic RAG baseline Self model Difference Overall accuracy 186 82.8% 90.9% +8.1 pp Hard questions 31 67.7% 87.1% +19.4 pp Tracking facts that changed over time 5 60.0% 100.0% +40.0 pp Point-in-time reasoning 6 33.3% 66.7% +33.3 pp Entity disambiguation 15 66.7% 86.7% +20.0 pp Multisource synthesis 73 87.7% 94.5% +6.8 pp Answered faithfully based on Corpus 13 100.0% 92.3% -7.7 pp Single-hop lookup 30 86.7% 83.3% -3.3 pp Citation coverage 186 92.5% 97.8% +5.4 pp Table 1. Evaluation metrics for the agentic RAG baseline and the self model in the example repo. Both configurations use NVIDIA Nemotron 3 Ultra Enforce the boundary at runtime This separation is enforced with NVIDIA NemoClaw and the NVIDIA OpenShell secure runtime for autonomous agents. NemoClaw integrates the example with NVIDIA OpenShell and manages its lifecycle, while NVIDIA OpenShell runs the agent in a sandbox and provides governance and policy enforcement for file system, process, and network access. For managed inference and MCP connections, credentials remain outside the sandbox. This is important because memory and retrieved content are inputs to the model, not trusted security policy. If the agent misinterprets that context—or follows malicious instructions—it still operates within operator-defined runtime boundaries. They limit the agent’s access and the potential impact of a failure. Get started building agent memory To adapt the memory design featured in this post for your own NemoClaw example, review the open source Memory-Driven Chief of Staff recipe and its design proposal in the NVIDIA/nemoclaw-community GitHub repo. The recipe packages the example as a deployable Hermes profile for NemoClaw: A structured memory schema A durable obligation ledger Bounded ranking logic User correction and audit paths Scheduled memory maintenance Synthetic messages and memory pages An offline walkthrough A suite of unit tests The sample people, organizations, projects, and messages are invented. Recorded model decisions stand in for inference in the offline walkthrough. The code then applies the ranking, correction, persistence, and validation behavior. The current recipe focuses on the memory foundation. It does not send messages or modify source systems. This scope allows you to examine the design without connecting a workplace account. Live connectors require separate handling for credentials, privacy, retention, and deletion. Learn more about NVIDIA NemoClaw and NVIDIA OpenShell .