Skip to content
BENIBIX Learn

Hermes Memory vs session_search

People think Hermes forgets anything not saved to memory. It does not. Memory is a few key facts loaded into every prompt, so it is small on purpose. session_search looks through every past chat on demand, at no model cost. Save the facts it needs every time, and let search find the rest.

AI🪽 Hermes Agent

Hermes MEMORY.md

The file of notes Hermes keeps for itself. It loads when each session starts.

MEMORY.md is the file where Hermes Agent keeps its own notes. They are facts about your machine, your projects and odd tool habits. It lives in ~/.hermes/memories and holds up to 2,200 characters by default. The agent writes it with the memory tool. When each session starts, its notes go into the system prompt as a frozen copy.

Example
Hermes learns that Mia's server runs Ubuntu with Docker. It saves that in MEMORY.md, so it never has to ask again.
Why it matters
A fact it needs on every run should not cost a question each time.
Full entry
AI🪽 Hermes Agent

session_search Tool

The Hermes tool that searches all your past chats. It costs no model calls.

session_search is the Hermes Agent tool that searches all your past chats. Every chat is stored in a SQLite database, ~/.hermes/state.db, that can search all its text. A search gives back the real messages, with no summary and nothing cut. It takes about 20 milliseconds and makes no model calls. Memory holds the key facts, and this finds the rest.

Example
Mia asks which database they picked last month. Hermes searches past sessions, finds the chat and quotes the answer.
Why it matters
Memory is small on purpose. Search lets the agent recall the rest without filling every prompt.
Full entry

In one line

Hermes MEMORY.md
The file of notes Hermes keeps for itself. It loads when each session starts.
session_search Tool
The Hermes tool that searches all your past chats. It costs no model calls.