explaingit

rednicv/hermes-improvements

Analysis updated 2026-05-18

★ 8PythonAudience · developerComplexity · 3/5LicenseSetup · moderate

TLDR

An add-on for Hermes Agent that uses vector search to inject only relevant memories into each prompt, cutting token costs by reusing prompt caching.

Mindmap

mindmap
  root((hermes-improvements))
    What it does
      Vector memory search
      Adaptive style learning
      Task routing
      Reasoning tracing
    Tech stack
      Python
      Embeddings
      TF-IDF fallback
    Use cases
      Cut API token costs
      Track reasoning steps
      Route tasks by complexity
    Audience
      Hermes Agent users
      Developers

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

What do people build with it?

USE CASE 1

Reduce API costs for a Hermes Agent setup by only loading relevant memories into each prompt.

USE CASE 2

Track how an AI agent's reasoning and source attribution evolved across a task.

USE CASE 3

Let an agent route simple and complex sub-tasks to different LLM models automatically.

What is it built with?

Pythonsentence-transformersNumPy

How does it compare?

rednicv/hermes-improvements100/awesome-machine-learningadam-s/car-diagnosis
Stars888
LanguagePythonPythonPython
Last pushed—2024-08-07—
Maintenance—Stale—
Setup difficultymoderateeasymoderate
Complexity3/51/53/5
Audiencedeveloperdeveloperresearcher

Figures from each repo's GitHub metadata at analysis time.

How do you get it running?

Difficulty · moderate Time to first run · 30min

Needs Hermes Agent already installed, embedding search needs extra RAM.

Use freely for any purpose, including commercial use, as long as you keep the copyright notice.

In plain English

Hermes Improvements Package is an add-on for Hermes Agent, a project by Nous Research. It does not replace Hermes Agent, instead it plugs into it and changes how memory and reasoning work under the hood. The main problem it solves is cost. Normally, Hermes Agent loads your entire memory file, meaning every past preference, rule, and detail you have taught it, into the system prompt on every single message. As that memory file grows, the prompt grows with it, and you end up paying for the same tokens over and over. This package changes that by adding a vector memory store. It uses embeddings, or a simpler TF-IDF method if you do not install the extra embedding library, to search your memories and pull out only the one to three that are relevant to what you just said. So if you mention a TV setup, it only injects your TV related memory, not your entire history. Because most memories stay out of the prompt, the system prompt itself stays stable, which lets providers like Google Gemini or Anthropic Claude reuse their prompt cache. In practice this can cut token costs by 50 to 80 percent on repeated turns. Beyond memory, the package adds a few other pieces. AdaptiveSoul learns your style preferences and corrections over time and keeps a log of the rules it has picked up. AdaptiveWorkflow looks at how complex a task is and decides whether to send it to a lightweight model or a more capable one. ReasoningTracer records the steps an answer took and where each piece of information came from. Installation involves copying the package files into your Hermes configuration folder and running an injector script that hooks into Hermes Agent's startup code without modifying it destructively. The injector checks whether the hook already exists before making changes, so it is safe to rerun after upgrades. It works fully on Linux and macOS, and falls back gracefully on Windows when file locking is not available. It requires Python 3.10 or newer, and at least 2 GB of RAM if you use the embedding based search. It is released under the MIT License, so it is free to use, modify, and share.

Copy-paste prompts

Prompt 1
Help me copy the hermes_improvements files into my ~/.hermes/improvements folder and run the inject hook script.
Prompt 2
Explain how VectorMemoryStore decides which memories to inject into a Hermes Agent prompt.
Prompt 3
Show me how to fall back to TF-IDF search if I do not want to install sentence-transformers.

Frequently asked questions

What is hermes-improvements?

An add-on for Hermes Agent that uses vector search to inject only relevant memories into each prompt, cutting token costs by reusing prompt caching.

What language is hermes-improvements written in?

Mainly Python. The stack also includes Python, sentence-transformers, NumPy.

What license does hermes-improvements use?

Use freely for any purpose, including commercial use, as long as you keep the copyright notice.

How hard is hermes-improvements to set up?

Setup difficulty is rated moderate, with roughly 30min to a first successful run.

Who is hermes-improvements for?

Mainly developer.

Open on GitHub → Explain another repo

This repo across BitVibe Labs

Verify against the repo before relying on details.

Alternatives & similar repos