Hermes Atlas
Memory & knowledge

Hermes Patches

Cyrene963/hermes-patches

Community patch set adding a Memory OS, semantic recall and pre-tool gating to Hermes Agent

In short

Hermes Patches is a community patch collection that overlays a layered Memory OS on the official Hermes Agent: structured long-term memory, semantic recall, pre-execution gating and a self-learning write loop. The README is written in Chinese.

What Hermes Patches does

The patches apply as an overlay on top of the official Hermes Agent and are described as verifiable, configurable and reversible. The Memory OS loop runs from perception to recall, pre-action gating, action and learning. Memory Graph provides structured long-term memory with 14 CRUD and search tools, version snapshots, aliases and term triggers, isolated per user with Postgres row-level security. Recall combines Hindsight vector embeddings and reranking with lexical recall over Memory Graph facts, and private memory ranks ahead of the shared public area.

A preflight gate checks tool arguments and blocks known failure patterns before execution, configurable in memory_policy.yaml. The write loop uses an LLM fact classifier with readback verification, writes only clean durable facts to a private namespace and pauses writes if the LLM endpoint is unreachable. Each capability is marked verified, partial or gated in docs/STATUS.md. The installer skips patches already merged upstream and targets Hermes Agent after v0.16.0 (v2026.6.5).

Key features

  • Memory Graph with 14 CRUD and search tools, version snapshots and aliases
  • Per-user isolation through Postgres row-level security, Hindsight banks and per-user MEMORY.md files
  • Semantic recall using Hindsight embeddings and reranking, with private memory ranked first
  • Preflight gate configurable in memory_policy.yaml that blocks known tool failure patterns
  • LLM fact-classifier write loop with readback verification that fails closed
  • Hybrid skill selector and a deep_research pipeline that writes evidence.json and manifest.json

When to use it

  • Giving Hermes Agent structured long-term memory that carries across messaging channels
  • Running a multi-user Hermes deployment with separate private memory per user
  • Stopping the agent from repeating tool calls that are known to fail

Who it is for: Self-hosting Hermes Agent users who want a Memory OS layer and are comfortable patching their ~/.hermes/hermes-agent checkout.

How it fits with Hermes Agent

Community patches applied as an overlay on the official Hermes Agent repository; they add only capabilities that are not yet merged upstream.

How to install Hermes Patches

These commands are copied from the project's README. Check the repository for the latest steps before you run them.

bash <(curl -sL https://raw.githubusercontent.com/Cyrene963/hermes-patches/main/install.sh)

Requirements: A real Hermes Agent repository at ~/.hermes/hermes-agent that contains toolsets.py, plus basic commands and Python dependencies; PostgreSQL (5432), Memory Graph (:8900) and Hindsight (:9177) are optional services.

Note: Capabilities are individually marked verified, partial or gated in docs/STATUS.md, and a running gateway must be restarted to load the changes.

FAQ

What is Hermes Patches?

Hermes Patches is a community collection of patches for Hermes Agent that adds a layered Memory OS. It covers structured long-term memory, semantic recall, tool pre-execution gating and an autonomous write loop.

How do I install Hermes Patches?

Run the install.sh script from the repository with bash and curl. It runs a preflight check first, is safe to repeat, and a running gateway must be restarted to load the Python changes.

Is Hermes Patches free and open source?

The repository has no license file, so default copyright applies. Check with the author before reusing the code.

Similar memory for Hermes Agent

All memory

Related guides: SOUL.md for Hermes Agent: what it is and how to write one · Run multiple Hermes agents with profiles