NVIDIA publishes a memory-driven agent recipe
NVIDIA has open-sourced a reference agent that keeps a structured, human-readable memory of a user's people and projects, and reports it beating an agentic RAG baseline 90.9% to 82.8% on the company's own 186-question benchmark.
NVIDIA's developer blog has published a reference implementation of a "memory-driven Chief of Staff" agent built on NemoClaw, its agent framework, and OpenShell, its sandboxed runtime. The example and its evaluation harness are in a public repo.
The design centers on what NVIDIA calls a self model: a human-readable knowledge layer of structured Markdown pages covering people, projects, priorities, goals, and recurring work patterns, with a schema defining indexing, cross-references, provenance, and growth limits. Judgments — whether an obligation needs attention, how it ranks, whether the user ignored it — live separately in a SQLite ledger with an append-only audit trail, so corrections persist across runs and repeated corrections roll up into an editable preference policy. Scheduled jobs refresh the memory as new activity arrives. An "intent gate" reserves the top priority tier for items tied to the user's stated priorities; the post's example ranks an urgent expense-policy attestation below a quieter request. OpenShell enforces file system, process, and network policy, keeping credentials outside the sandbox on the premise that memory and retrieved content are model inputs, not security policy.
On NVIDIA's own 186-question Agent Memory Benchmark, the self model scored 90.9% overall against 82.8% for an agentic RAG baseline, with the largest gains on hard questions (+19.4 pp) and facts that changed over time (5 questions, +40 pp). It lost ground on single-hop lookup (-3.3 pp) and corpus-faithful answering (-7.7 pp). The benchmark, the baseline, and the results are all NVIDIA's; no external replication is offered.