DWF Ventures published an analysis on June 5 arguing that most AI agents remain “fundamentally stateless,” losing all context and learned behavior after each session. The firm highlighted Nous Research’s open-source Hermes framework as a structural answer to the problem, according to WuBlockchain, which first reported the note.
Hermes, released in February 2026 under an MIT license, introduces persistent memory that retains user interactions, preferences, and project context across sessions. An automated Skills system writes reusable skill documents whenever the agent solves a complex problem, making those solutions permanently available for future tasks. A self-improvement loop continuously refines what the agent knows, compounding capability rather than resetting every cycle.
What Statelessness Actually Costs
The stateless architecture that dominates current AI agents exists because it is cheap and avoids storing sensitive data. That tradeoff worked for simple alert bots and single-command automation. It breaks down when agents manage multi-step workflows, interpret financial positions, or maintain long-running projects that span weeks.
DWF’s note explicitly compared Hermes to Claude Code and OpenAI Codex, according to Blockchain Reporter. The argument: those tools generate strong output in the moment but cannot compound knowledge across sessions. An agent that produces a perfect code audit on Monday forgets the entire codebase by Tuesday.
How Hermes Retains State
Hermes stores all memory locally in ~/.hermes/ on the user’s machine. Cross-session recall uses FTS5 search with LLM summarization, and user modeling runs through a dialectic profiling system called Honcho, according to MarkTechPost. Skills follow the agentskills.io open standard, making them portable and shareable across installations.
The framework bakes in credential isolation, secret redaction, and automatic key rotation to separate access tokens from the agent’s reasoning layer. That security architecture matters because stateful agents holding user credentials become high-value targets.
Hermes integrates with Nous Research’s decentralized Psyche training network, which distributes model refinement across a node structure rather than relying on a single server. The official site lists 40+ built-in skills, five chat platform integrations (Telegram, Discord, Slack, WhatsApp, Signal), and execution backends spanning local, Docker, SSH, Singularity, and Modal.
Desktop and Adoption Momentum
Nous Research shipped Hermes Desktop in public preview on June 2, adding a native GUI for macOS, Windows, and Linux. The desktop app shares configuration, API keys, sessions, and memory with the CLI and messaging gateways. Sessions started on desktop resume in the terminal and vice versa.
The desktop release was first demoed at Jensen Huang’s GTC keynote. It runs against the same Hermes Agent v0.15.2 core, not a fork.
The Open-Source Tradeoff
Hermes’ MIT license invites community auditing and adaptation, which could accelerate uptake in developer tooling and automation workflows. Maintaining a security edge while staying fully open-source against well-funded, closed-source competitors (Anthropic, OpenAI) is the central tension. DWF’s spotlight signals that venture capital is evaluating agent infrastructure on architectural durability, not just user acquisition numbers. Whether developers commit to running a single, long-running agent environment at scale remains the open question.