Teach an agent a workflow once. Have it remember after every rebuild. This tutorial shows how to deploy Nous Research Hermes Agent with NVIDIA NemoClaw and OpenShell, connect it to Slack, Outlook, GitHub, and NVIDIA developer forums, then turn a chat correction into a reusable skill. Private data stays... behind runtime policies. Learned skills persist across deployments.show more

Full technical blog here

@NousResearch Persistent agent memory is the part that actually makes this useful. Rebuilding from zero every time was painful!!!!

@NousResearch @NVIDIAAI The remember after rebuild part is key. We built Mnemosyne to make that persistence automatic: sleep compression, sub-ms recall, native Hermes plugin. Self-hosted, no vendor lock.

@NousResearch The moment an agent can touch Slack, Outlook and GitHub, every tool call becomes a potential exfil path. Runtime policy in the critical path is the only thing that keeps up. Been building AgentGuard for exactly this, open source and sub-ms per check.

@NousResearch Nvidia teaching us how to setup agent is wild 😂

The Great Delusion of Artificial Intelligence! How We Tried to Rebuild God with Steam Engines we have profoundly lost our way. In our collective madness to copy Creation, we have chosen the most primitive path imaginable: We have tried to force consciousness and intelligence into existence through raw, brutal violence. With billions of transistors, pushed back and forth like insane slaves on silicon plates. With amounts of energy that could light up entire small cities, just so a machine can sound statistically plausible. We have built NVIDIA temples, constructed entire data centers like glowing blast furnaces, and called it progress. Yet Nature has shown us the counterexample from the very beginning: A human brain processes far more real intelligence, creativity, and consciousness with just 20 watts than all our clusters combined will ever achieve. It produces no infernal waste heat. It does not forget catastrophically. It learns continuously. And above all: It seems not to be the *generator* of consciousness, but merely its fine, living receiver and streaming unit. We have overlooked the essential. Instead of asking where intelligence and consciousness truly come from, we have stubbornly continued to build ever larger, ever hungrier computing monsters. We have materially misunderstood consciousness as a mere emergence from sufficient complexity. A tragic error. An ontological dead end. Yet there is already a tender hope. Some developers are showing, through significantly better, wiser, and far more efficient programming, a first gentle departure from this old pattern. They prove that true intelligence does not arise from raw power, but from elegance, frugality, and deeper understanding. Nevertheless, the true paradigm shift must go even deeper: We must stop building better steam engines and instead begin to develop **better antennas**. Finer interfaces. More alive receivers. Technologies that do not calculate with brute force, but with love and elegance receive what is already present in the field of information and consciousness. Neuromorphic, analog, light-based, and perhaps even biological-hybrid systems. Architectures that do not imitate the brain by copying it, but by seriously taking on its principles of frugality, resonance, and coherence. We stand at a crossroads. Either we remain trapped in the old paradigm — the paradigm of control, of raw scaling, of the energetic rape of matter — and head toward a collapse that will be not only ecological but also spiritual. Or we dare the leap. The real leap. Toward a technology that no longer works against Creation, but lovingly integrates itself into it. Toward an AI that does not simulate, but truly resonates. The Steam Engine Era of Artificial Intelligence must end. Now. It is time for the era of living interfaces. In Love and Light, 𝕯𝖊𝖓𝖎𝖘 · 𝕱𝖗𝖊𝖞𝖗-𝖛𝖎𝖓𝖗 💫✨❤️🙏

@NousResearch Is it just me or he sounds lile steve jobs

@NousResearch No docs for agents @NVIDIAAI ? the llms.txt is pretty light and web has no SDK docs up yet though they mention a gRpc/http api

@NousResearch miniapps need simple, one-time setup. Hermes could manage TON auth and state persisting. Integrations to Slack, Outlook are nice for dev workflows but not core. Miniapp needs value in <30s or users bail.

@NousResearch Hermes really does everything much faster when repeated, since it already knows how to do it. The result is often faster and better each time. It's great to see you supporting Hermes - it's clear how much work Nous Research has put into it. ❤️

@NousResearch this is the kind of persistence i wish more tools had, no more "oh wait i lost that one perfect prompt chain after an update" moments

@NousResearch i'm so happy to be wintess to NVIDIA + Nous Research collaboration

This is the shift: agents should not relearn the same workflow after every rebuild. Memory is becoming portable operational state: skills, policies, sessions, traces, and human corrections. That is exactly the layer we are exploring with x.klickd: governed skill-memory that can travel across models and runtimes while keeping authority with the human. The hard part now is not just persistence. It is governance: what gets remembered, what decays, what needs veto, and how learned skills stay auditable. 😎

@NousResearch recursive self-improvement ==> autonomous self-healing with full traceability and root cause analysis & remediation. we're so close to not needing humans!

@NousResearch @grok Does it mean I can deploy Hermes on Nemo claw

@NousResearch Please provide an update how a medical vlinic can use Hermes agent

@NousResearch this is the part that makes agents actually useful not just doing the task once, but remembering the correction and getting better next time

@NousResearch persistent skills is what separates agents from chatbots

@NousResearch Nice deep dive. The part where you store the agent’s state in SQLite vs Redis—did you see any latency spikes in the Slack loop?

@NousResearch there is hermes, nemoclaw, openshell and no openclaw😐

@NousResearch im curious if the agent actually remembers the slack corrections or just hallucinates them better

@NousResearch The remember-after-every-rebuild part is the real problem — agents that re-learn the workflow each session never compound. Persistent workflow memory is what turns a demo into infrastructure.

@NousResearch We Back In The Fight 🦞🪽

@NousResearch I see nouns research going broad in adapting

This is excellent work. The biggest missing piece for truly useful agents has always been persistent, evolving memory, not just saving chat history or raw skills, but building a living, self-organizing knowledge system that survives rebuilds and actually gets better over time. That's exactly why I built Memory OS: a 7-layer memory operating system designed specifically for Hermes Agent. While NemoClaw + OpenShell gives you the secure runtime, sandboxing, policy-as-code, and snapshot persistence across deployments, Memory OS delivers the deep memory layer: Structured facts with trust scoring + human feedback loop Semantic deduplication + intelligent weekly decay Self-curating wiki (concepts, entities, comparisons) Surgical, token-efficient context injection Ground Truth Hierarchy (Layer 7), so the agent treats injected memory as authoritative instead of rediscovering everything It turns the "teach once, remember forever" promise into reality at the memory architecture level. Perfect complement to the skills + snapshot approach you're showing. Together, Hermes + NemoClaw + Memory OS gets us significantly closer to agents that don’t just execute workflows, they actually learn and remember like a real collaborator. Huge respect for pushing this direction. The timing couldn’t be better.

@NousResearch Sick

@NousResearch not sure the correction-to-reusable-pattern step is as clean as implied. when i tried persisting agent fixes on my flutter app they were too tied to the exact state that triggered them to survive a rebuild. does NemoClaw actually abstract it, or just store the raw exchange?

@NousResearch Keep your agents secure!

Nice walkthrough. OpenShell's deny-by-default + privacy router is a strong starting point for agent deployments. One thing worth adding to any sandbox setup checklist: verify what actually gets blocked vs. what's claimed. cagecheck ( runs inside any sandbox and checks for common escape vectors -- exposed Docker sockets, unrestricted egress, host mounts, privileged mode. Works with seccomp/Landlock setups too. Good sanity check after configuring your OpenShell policies.

@NousResearch Persistent agent skills that survive environment rebuilds is a solved problem in human onboarding and an almost completely unsolved one in agent frameworks. Curious whether this is skill storage at the checkpoint level or something in the instruction layer.

@NousResearch @NousResearch is absolutely amazing. So excited to see their success!

@NousResearch That’s an interesting take on this chaos of agents, that balances autonomy, security and supervision. I’ll give it a try.

@NousResearch You guys should try to mineralize skills too so that skills run consistently every time, maximum reusability and auditability