This is a submission for the Hermes Agent Challenge
A very good evening everyone!
Hermes agent latest version is new in town and I can't keep it to myself that I am very glad that I could complete the build challenge yesterday by showcasing personal AI newspaper in which the user receives personalised news according to their cup of tea everyday.
Well! Let's not go there. It's been more than 5 months learning Agentic AI and I barely scratched the surface with making useful projects in this particular domain.
If you know you know that Hermes agent is the self improving AI agent. Plus its session search is 4,500x faster this time and the agent is absolutely free. It is having built-in learning loop yes I am discussing about GEPA memory.
What is GEPA
Generate → Evaluate → Prune → Accumulate
Generate — an agent attempt to do some task and learns skill based on it
Evaluate — assigns grades to the skill whether it is helpful or not
Prune — it deletes the unnecessary skill which won't be used likely
Accumulate — it saves the skills for the next runs
Basically it is a learning loop which learns as a human does.After each run the agent gets smarter because it retains the memory of the job done in the past.
The Memory System
It does not only utilise skill memory but also uses FTS5 full-text search on past discourse which means the context can be found at ease from the long corpus of given text.
It works as a curator agent running in the background and uses GEPA to make the context stronger and relevant. In other words I should say The agent itself discards the memory which is considered obsolete.
Hermes agent's Core Architecture
Interface layer — user communicates with the agent (CLI, Telegram, WhatsApp etc.) supports 20+ platforms.
Agent core — LLM, planner, and tool dispatcher reside in the agent core
Tool layer — 40+ built-in tools such as web search, browser automation, vision, file system, code execution which the agent uses
Memory system — GEPA loop
Output layer — final result in the form of text, files, emails and code
It costs nothing when you keep it idle and provides MCP support as well.You can run it on a $5 VPS, a GPU cluster, or serverless infrastructure through Daytona or Modal. Even if you are not working for example using phone for some research task your work will be finished after you come back without keeping local machine running. That's the power of Hermes Agent
Courtesy: Claude for Architecture Diagram
Top comments (1)
"Grows with you" is the aspiration, and the honest engineering reality is that the base model doesn't grow - so all the growth has to come from the system around it: accumulated preferences, your patterns, corrections it remembers, context that compounds. The agent feels like it's learning because the harness is getting better at feeding it the right context, not because the weights changed. Naming that clearly matters, because it tells you where to invest (the memory/context layer, not waiting for a smarter model).
The design tension I keep hitting with "grows with you" systems: more accumulated memory eventually becomes noise, and a preference you had 3 months ago can actively mislead today. So growth needs forgetting/curation as much as accumulation. I wrestle with this exact balance in Moonshift (a multi-agent pipeline shipping a prompt to a real SaaS) - durable context that compounds, but scoped/curated so it stays sharp. Nice piece - how are you handling the "remembered preference is now stale" problem? That's the part that decides whether growth helps or slowly poisons.