Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Self-host Honcho memory layer for Hermes Agent — OpenRouter + Venice, no code changes
| Date | Stars |
|---|---|
| 2026-07-31 | 352 |
| 2026-08-03 | 354 |
| 2026-08-06 | 354 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# Self-Hosted Honcho for Hermes Agent
Self-host [Honcho](https://github.com/plastic-labs/honcho) (Plastic Labs' memory layer) on your own server instead of using their cloud. Works with [Hermes Agent](https://github.com/NousResearch/hermes-agent) out of the box.
**No fork required** — just 3 config files on top of upstream Honcho.
## Background: Hermes L4 Memory
Hermes Agent has a 4-layer memory system. The cross-session memory layer is powered by [Honcho](https://github.com/plastic-labs/honcho), which builds a deepening model of the user across conversations — extracting observations, recalling context, and consolidating memories over time.
By default, Hermes uses Plastic Labs' managed cloud ([honcho.dev](https://honcho.dev)) + their [Neuromancer](https://plasticlabs.ai/neuromancer) models. This works out of the box but means your conversation data and user profile live on their servers.
### What are Neuromancer models?
[Neuromancer XR](https://blog.plasticlabs.ai/research/Introducing-Neuromancer-XR) is a specialized 8B model fine-tuned from Qwen3-8B specifically for extracting logical conclusions from conversations. Unlike general-purpose LLMs which are optimized for plausible text generation, Neuromancer is trained on ~10,000 curated social reasoning traces to follow formal logic — extracting both explicit facts ("user said they like Python") and deductive conclusions ("user is likely a developer").
It scores 86.9% on the LoCoMo memory benchmark vs. 69.6% for base Qwen3-8B and 80.0% for Claude 4 Sonnet.
**Tradeoff of not using it:** General-purpose models work well for observation extraction and memory recall — Honcho's prompts and tool-calling pipeline compensate for much of the gap. You may get slightly less precise deductive reasoning, but capable models (GLM-5, Grok 4.1) with strong function calling largely close the difference. The main advantage of self-hosting is data sovereignty, not matching Neuromancer's exact reasoning quality.
## Deployment Options
| Option | Privacy | Data location | LLM for memory | Setup | Cost |
|--------|---------|--------------|----------------|-------|------|
| **Managed cloud** (default) | Low — data + inference on 3rd party | Plastic Labs servers | Neuromancer (Plastic Labs) | None — built into Hermes | Free tier / paid |
| **Self-hosted + API** (this repo) | Medium — data on your machine, inference via API | Your machine | Any OpenAI-compatible API | ~3 minutes | API usage only |
| **Self-hosted + local model** | High — nothing leaves your network | Your machine | Local LLM (Ollama, vLLM) | More setup | Hardware only |
**Managed cloud** — Zero setup. Best for getting started. Your data is on Plastic Labs' infrastructure.
**Self-hosted + API** — This repo. Your data stays on your machine. LLM calls go to a cloud API for inference only — the provider sees request content but doesn't store your memory data. Best balance of privacy and capability.
**Self-hosted + local model** — Maximum privacy. No data leaves your network. Requires a GPU or capable CPU on your LAN running an inference server (Ollama, vLLM, llama.cpp). Set `LLM_VLLM_BASE_URL` to your local server. Trade-off: smaller models may produce lower quality observations and reasoning than cloud APIs.
## What this does
- Runs Honcho's full memory stack (API, Deriver, PostgreSQL, Redis) on your machine
- Routes LLM calls through any OpenAI-compatible provider (primary + backup)
- All your data stays on your machine — no third-party cloud storage
- Works with OpenRouter, Venice, Routstr, Together, Ollama, or any other provider
## Architecture
```
Hermes Agent ──► localhost:8000 (self-hosted Honcho API)
│
├── PostgreSQL + pgvector (your machine)
├── Redis cache (your machine)
│
└── Deriver/Dialectic/Dream workers
│
├── Primary LLM provider (any OpenAIExcerpt of 16,446 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:1a4a4151b47c728a, llm:Repository description: 'Self-host Honcho memory layer for Hermes Agent — OpenRouter + Venice, no code changes' (language: Shell).
matched fp:1a4a4151b47c728a, llm:Repository description: 'Self-host Honcho memory layer for Hermes Agent — OpenRouter + Venice, no code changes' (language: Shell).