Hermes Agent Skills and Memory

The first time Hermes recovers a memory you forgot you mentioned, the difference between an agent and a chat product becomes obvious. The chat product has no past. The agent does.

The single feature that most defines Hermes Agent against the closed-source agent products is persistent memory. Other agents reset when the conversation ends. Hermes keeps state across hours, days, and weeks. This guide explains what gets stored, how skills emerge from that storage, and what you should know about controlling the memory layer.

How does memory actually work in Hermes Agent?

Three storage layers, each with a different lifecycle and purpose.

• Short-term memory. The current conversation. Same as any chat product. Held in the model’s context window and discarded when the session ends.
• Long-term memory. Selected facts, decisions, and preferences extracted from past conversations. Stored in a local database on your server. Hermes consults it when relevant context comes up in future conversations.
• Skills library. Successful patterns the agent has learned to repeat. If you asked Hermes to draft a weekly board update three times in a row with similar structure, it promotes that pattern to a reusable skill. Next week you ask for “the board update,” it pulls the skill.

The three layers work together. Short-term holds what is happening now. Long-term holds what was decided last month. Skills hold the playbook for repeating successful work.

What gets stored in long-term memory?

Not everything. Hermes is selective by design because storing every conversation in raw form would bloat the database and degrade retrieval quality. The agent uses an extraction model to pull out the high-value pieces.

• User preferences (“I prefer concise answers, no bullet points unless asked”)
• Project context (“I am working on the Q3 board update”)
• Decisions (“we decided to launch the feature in October”)
• Recurring tasks (“every Monday morning, do X”)
• People and relationships (“Alex is the PM on the design team”)
• Tool credentials and connection details (encrypted)

The agent does not store full conversation transcripts by default. If you need that for audit or compliance, the option is configurable in the Hermes settings. Default off because most users do not need it and storage adds up.

How do skills get created?

A skill is a learned pattern the agent can re-execute. The lifecycle has three stages.

1. Observation. You do a task with the agent. It works. The agent logs the task structure, the inputs, the steps it took, the output you accepted.
2. Pattern detection. After you do similar tasks a few times (the threshold is configurable, default 3), the agent recognizes the pattern.
3. Skill promotion. The agent generates a named skill (e.g., “draft-board-update”) with parameterized inputs. Next time you describe the task in plain English, the skill activates automatically.

You can also write skills manually if you do not want to wait for the agent to learn them. The skill format is a YAML file in ~/.hermes/skills/. For terms like YAML, agents, or context window that come up here, the AI Glossary has plain-English versions. The format is the same the auto-learned ones use, so you can edit auto-learned skills the same way.

How much control do I have over what is remembered?

You control three things explicitly.

• What to remember. Tell Hermes “remember that X” and it adds X to long-term memory immediately. Tell it “forget X” and it removes the entry. The agent confirms each action.
• What to forget on a schedule. Configure retention rules per memory category. By default, factual decisions persist indefinitely; project context expires when the project closes; conversation summaries roll off after 90 days. Adjust to taste.
• What to share across users. If multiple people use the same Hermes instance (e.g., a team Slack bot), you can scope memory per user or share it across the team. Privacy-sensitive items default to per-user.

For the broader security framework around what the agent should and should not have access to, see the Zero Trust for AI Agents guide.

When will I feel the memory working?

The first time the agent surprises you with a recovered memory. Common moments:

• You ask a question about a project and the agent says “by the way, you mentioned three weeks ago that the deadline moved to October 15; does that still apply?”
• You start a task and the agent says “I have done this kind of task before; should I use the same template as last time?”
• You log in after a vacation and the agent says “while you were away, here are the three things from your scheduled tasks that completed successfully and the one that needs your attention.”

This is the moment the difference between Hermes and a stateless chat product becomes obvious. The chat product cannot do any of this because it has no past to draw on. Hermes does because it does.

How do I inspect, edit, or back up the memory?

Three commands handle most of it.

• hermes memory list — shows what is in long-term memory, organized by category.
• hermes memory remove [item] — deletes a specific memory entry.
• hermes memory export — dumps everything to a JSON file for backup or migration. hermes memory import restores from a backup.

For skills, the equivalent commands are hermes skills list, hermes skills remove [name], and direct edits to the YAML files in ~/.hermes/skills/.

Frequently asked questions

Does memory work across chat channels?

Yes. If you talk to Hermes in Slack on Monday and Discord on Tuesday, both channels share the same memory and skills.

Can I see what Hermes remembers about me?

Yes. hermes memory list shows everything. Audit it whenever you want. Remove anything that should not be there.

What happens to memory if I uninstall Hermes?

Deleting ~/.hermes/ removes all stored memory and skills. The data lives on your server only; nothing was sent to a vendor.

Is memory encrypted at rest?

Credentials and explicitly-flagged sensitive fields are encrypted. General conversation memory is stored in plaintext in the local database. Run on encrypted disk if you need full-disk protection.

Can skills be shared between Hermes instances?

Yes. Skills are YAML files; you can copy them between installs. Useful for setting up a fresh agent with the skills you trained on another.

Sources

• Hermes Agent docs
• Nous Research
• Nous Research on Hugging Face
• Nous Research blog

You might also like

• Hermes Agent Explained
• How to Install Hermes Agent
• Hermes Agent Discord Quickstart
• Zero Trust for AI Agents
• Hermes Agent vs Claude Code
• AI Glossary