Hermes Agent: the open-source agent that turns self-improvement into a closed learning loop

What it is

Hermes Agent is the self-improving agent from Nous Research. Its differentiating claim is not better tool calling but a closed learning loop: it creates skills from experience, improves those skills while using them, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions. It has more than 248,000 GitHub stars, installs with a single curl script, and offers a native PowerShell path on Windows.

Its assumptions about where it runs are wide: a $5 VPS, a GPU cluster, or serverless infrastructure that costs nearly nothing when idle. It is not tethered to your laptop - you can talk to it from Telegram while it works on a cloud VM. The model side is not locked either: Nous Portal, OpenRouter, OpenAI, and your own endpoint all work, and hermes model switches between them with no code changes.

The learning loop, mechanism by mechanism

The README decomposes the loop into five mechanisms, and they are worth separating because together they are the entire substance behind the phrase "self-improving". First, agent-curated memory with periodic nudges: the system reminds the agent at intervals to write down what is worth keeping, rather than compressing only when context overflows. Second, autonomous skill creation after complex tasks - the procedure that just worked gets deposited as a reusable skill. Third, skills that self-improve during use: the next time a skill is invoked, it revises itself against that run's experience. Fourth, FTS5 session search paired with LLM summarization for cross-session recall. FTS5 is SQLite's full-text extension, which means conversation history lives in local SQLite, retrieval is lexical, and recall hands results to an LLM for summarization. Fifth, Honcho dialectic user modeling, accumulating a picture of who you are across sessions.

Mechanisms two and three carry the real engineering weight, because they turn memory from text into an executable asset: a skill is something that gets run, not something that gets read. The fourth mechanism choosing FTS5 over vector retrieval is a conservative but explainable decision - lexical search has no embedding-model dependency, runs locally, and produces results you can trace.

It is also compatible with the agentskills.io open standard, so skills it produces can be consumed by other agent runtimes. That is where it differs from the DeepSeek Harness route of "plugins come from the ecosystem": Hermes skills mostly grow out of the agent's own operation.

Shape: one TUI, seven terminal backends, six chat platforms

The interaction surface is a genuine terminal interface: multiline editing, slash-command autocomplete, conversation history, interrupt-and-redirect, and streaming tool output. Interrupt-and-redirect matters in long tasks - you can change direction mid-run instead of killing it and starting over.

Reach covers Telegram, Discord, Slack, WhatsApp, Signal, and CLI from a single gateway process, with voice-memo transcription and cross-platform conversation continuity. Execution has seven terminal backends: local, Docker, SSH, Singularity, Modal, Daytona, and Vercel Sandbox. Daytona and Modal provide serverless persistence - the environment hibernates when idle and wakes on demand, so between-session cost is near zero. Singularity in that list says the project accounts for HPC environments, the common scheduling shape in academic and research clusters.

Parallelism works by spawning isolated subagents for separate workstreams, and by writing Python scripts that call tools over RPC - which the README describes as collapsing multi-step pipelines into zero-context-cost turns. The claim holds up: when a script calls tools, intermediate results need not return to the main context, only the final one does.

The research-facing side

Nous Research is a model company, so Hermes doubles as a data-production apparatus: batch trajectory generation, plus trajectory compression aimed at training the next generation of tool-calling models. That distinguishes it from purely application-shaped harnesses - its conversations are not only doing your work, they are also training corpus. Users should know that.

Boundaries, and what you have to judge yourself

The installer handles uv, Python 3.11, Node.js, ripgrep, and ffmpeg for you, and on Windows adds a portable Git Bash (MinGit, unpacked to %LOCALAPPDATA%\hermes\git, no admin required, isolated from any system Git). The README devotes a whole section to antivirus software flagging uv.exe as malware, and gives a complete GitHub CLI attestation sequence to verify your copy of uv is authentic. That documentation is responsible, but it also tells you something: the install chain downloads and executes unsigned Rust binaries, so a strict security environment needs an allowlisting process.

"Self-improving" is a falsifiable claim, and we have not run a controlled experiment testing whether skill improvement actually raises success rates on later tasks, or merely raises the number of skills. The stars and the narrative are both strong; the capability level needs reproduction. We grade this entry as needing reproduction.