HHermes Explained

Memory

Long-Term Memory

Curated information intended to remain useful across sessions and days.

Importance 5/55 backlinks1 source refs

Long-Term Memory

What it is

Curated information intended to remain useful across sessions and days.

In plain terms

Long-term memory is the small set of facts Hermes decides belong in the permanent record, similar to how a contact card keeps someone’s current address and preferences up to date across years of knowing them, rather than saving a transcript of every conversation you’ve ever had with them. It’s meant to still matter in a month, not just for the rest of today’s conversation.

Why it matters

Not every detail from a conversation earns a place in long-term memory — only the subset curated specifically to stay relevant across sessions and days, unlike the full raw log session state keeps. That distinction is what lets a daily priority ritual reference a goal set a week earlier, or a research briefing build on an earlier run’s findings, without a user re-explaining context each time. Session state dies with its session; long-term memory is meant to outlive it.

How it works

This layer is exactly what the built-in memory files hold, or what an external provider’s persistent store holds if one is configured: MEMORY.md for the agent’s own operating notes and USER.md for facts about the person, each capped by a character limit — 2,200 and 1,375 characters by default — so the store can’t balloon into something too bulky to inject whole into every prompt (tools/memory_tool.py). Because both files load once and freeze into the system prompt at session start, a fact saved here is visible to the model on literally the next message it sees in a later session, without being searched for. When either file nears its limit, adding something new is rejected until stale entries get trimmed or replaced, keeping the store deliberately small and current.

A concrete example

Across a week of sessions, a founder tells Hermes to always draft investor emails in a particular tone, then later mentions the company just closed a new office lease.

  1. The tone preference goes into USER.md as a standing instruction.
  2. The lease detail goes into MEMORY.md as a fact about the current situation. The next month, asked to draft a partner update, Hermes already has both facts sitting in context — no re-explaining needed, because they were curated into the permanent store instead of just existing somewhere inside an old chat log.

How it connects

This note belongs to the Memory family. Its closest neighboring concepts are: Memory Provider, Session Search, Agent Runtime, Personal OS.

Source evidence

  • vendor/hermes-agent/agent/memory_manager.py:505-515#MemoryManager.prefetch_all — Verified Source Map evidence for MemoryManager. MemoryManager calls prefetch on each registered provider in order, skips empty results, isolates individual provider failures, and merges the remaining text.

Workflows

  • No first-pass workflow mapping yet; this concept still appears in the vault graph through articles and source evidence.
  • Hermes Agent for Beginners: A Step-by-Step Route from Empty Server to Useful Automation
  • Why Hermes Feels Different: Memory, Self-Improving Skills, and an Agent That Keeps Working
  • Inside Hermes: How the Agent Loop, Context, Memory, Gateways, and Cron Jobs Work Together
  • The Ultimate Beginner’s Guide to Hermes: From First Installation to an AI Operations Team