The best Hermes Agent memory setup is the built-in memory plus session search, with skills for anything you repeat. That combination ships with Hermes, costs nothing extra, and covers what most people need. External memory providers are worth adding later, once you know exactly what the built-in setup can't do for you.
📺 Watch: Hermes Obsidian Just GOT BETTER!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
This guide ranks the ten memory options worth knowing, explains how Hermes memory actually works, and shows you how to set it up without filling your agent's head with junk. Everything here comes from the Hermes docs, so where a provider is listed, read its own README before you commit.
How Hermes Agent memory works
Per the Hermes docs, built-in memory is two plain files in ~/.hermes/memories/.
- MEMORY.md holds the agent's own notes. It's capped at about 2,200 characters, roughly 800 tokens.
- USER.md holds your profile. It's capped at about 1,375 characters, roughly 500 tokens.
Both files load once, at the start of a session, as a frozen snapshot. If the agent saves something mid-session, it goes to disk straight away, but it won't show up in the agent's context until the next session.
The memory tool can add, replace and remove entries. It can't read them, because the snapshot is already in context. If a write would push a file over its limit, Hermes returns an error instead of quietly dropping old entries. That forces the agent to consolidate. Duplicate entries are rejected and every entry is security-scanned. Memory is on by default.
That's small on purpose. A few thousand characters is enough for who you are, how you like things done, and the handful of facts that matter every single session. It's not a place for whole projects.
If you want the full walkthrough of building agents that keep context across weeks of work, the AI Profit Boardroom has Hermes courses plus my Agent OS, which bundles memory, skills and tools in one place.
How we ranked the best Hermes Agent memory options
Three questions decided the order:
- Does it ship with Hermes? Built-in options win because there's nothing to install, nothing to pay for and nothing to break.
- Does it solve a real limit? The built-in snapshot is small and per profile. Options that fix those two gaps rank higher.
- How much do the docs say about it? Providers in the Plugin Catalog are listed, but the docs don't describe each one in depth, so they sit lower until you've tested them yourself.
1. Built-in memory (MEMORY.md + USER.md)
This is the foundation and the reason Hermes feels like it knows you. Your preferences, your stack, your business and your tone live in USER.md. The agent's working notes live in MEMORY.md. Both load every session automatically.
External providers run alongside built-in memory, per the docs. They don't replace it. So whatever else you add, this stays switched on.
2. Session search
Session search is the second half of the built-in setup. The session_search tool searches your past conversations, which Hermes stores in SQLite at ~/.hermes/state.db, using full-text search. It returns the actual stored messages, not an AI summary of them.
That's why it pairs so well with the small memory files. MEMORY.md holds the facts that matter every time. Session search digs up the detail from three weeks ago when you need it. To browse past sessions yourself, run hermes sessions list.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
3. Skills
Skills aren't technically memory, but the docs recommend saving recurring workflows as skills rather than memory. Skills don't use the character budget, so your memory files stay free for facts.
If you catch yourself telling the agent "every time you write a post, do these six steps," that's a skill, not a memory entry. For the ones worth installing, see my list of the best Hermes Agent skills.
4. Obsidian vault as a second brain
I've covered Hermes plus Obsidian in several videos. The idea is simple: you point the agent at an Obsidian vault and it reads and writes notes there. Your knowledge base becomes something the agent can search and update, and you can read it yourself in Obsidian.
It's not a memory provider in the plugin sense. It's a folder of notes the agent uses through its file tools. That's exactly why it works so well for long-form knowledge that would never fit in 2,200 characters. Here's the full setup in my Hermes second brain guide.
5. Honcho
Honcho is in the Hermes Plugin Catalog. Per the docs, external providers like this add things the built-in files can't, such as knowledge graphs, semantic search, automatic fact extraction and cross-session user modelling. Which of those Honcho does, and how, is in its own README. Install it with hermes plugins install honcho, then test it on real work before you trust it.
6. Mem0
Mem0 is also in the Plugin Catalog. Same rule applies: it runs alongside built-in memory, and you should check its README for exactly what it stores and where. Install with hermes plugins install mem0.
📺 Watch: Hermes AI Agents Just Went Portable
7. Supermemory
Supermemory is another catalog provider. If you're comparing providers, set up one at a time, run the same week of tasks, and see which one actually brings back the right context. Install with hermes plugins install supermemory.
8. Hindsight
Hindsight is in the catalog too. Install it with hermes plugins install hindsight, then read its README before you rely on it for anything client-facing.
9. Holographic, RetainDB and ByteRover (bundled)
These three come bundled with Hermes, per the docs, so there's no catalog install step. You pick one during hermes memory setup. They sit at number nine because the docs list them without much detail on each, so you'll want to test them side by side yourself.
10. OpenViking
OpenViking rounds out the catalog providers. Install with hermes plugins install openviking. Like the rest, it adds to built-in memory rather than replacing it.
Want help picking the right memory setup for your actual business? Book a free AI strategy session and we'll map it out.
Ranking table
| # | Memory option | Type | Install | Best for |
|---|---|---|---|---|
| 1 | Built-in memory | Built in | On by default | Facts and preferences every session |
| 2 | Session search | Built in | On by default | Finding exact past messages |
| 3 | Skills | Built in | Save as a skill | Recurring workflows |
| 4 | Obsidian vault | Notes folder | Point the agent at the vault | Long-form knowledge base |
| 5 | Honcho | Catalog provider | hermes plugins install honcho | Testing an external provider |
| 6 | Mem0 | Catalog provider | hermes plugins install mem0 | Testing an external provider |
| 7 | Supermemory | Catalog provider | hermes plugins install supermemory | Testing an external provider |
| 8 | Hindsight | Catalog provider | hermes plugins install hindsight | Testing an external provider |
| 9 | Holographic / RetainDB / ByteRover | Bundled provider | Pick in hermes memory setup | External memory with no catalog install |
| 10 | OpenViking | Catalog provider | hermes plugins install openviking | Testing an external provider |
How to set up Hermes memory
Built-in memory and session search are already on. For an external provider:
- Install it if it's a catalog provider, for example hermes plugins install mem0.
- Run hermes memory setup and choose the provider.
- Run hermes memory status to check it's connected.
- Start a new session and give it real work.
If you want control over what gets saved, turn on write_approval in your config. It's off by default. With it on, Hermes asks before saving memories, and you can review staged writes with /memory pending. That's worth doing for the first couple of weeks, so you see exactly what the agent thinks is worth remembering.
If you're also adding other plugins, my list of the best Hermes Agent plugins covers what else is worth installing.
Tips from the Hermes docs
- Consolidate at about 80% full. Don't wait for the write error. Merge and trim entries before you hit the limit.
- Run /new at natural breaks. New memories only load next session, so starting fresh brings them in and keeps your context short.
- One agent per Hermes home. Memory is per profile. Don't point two agents at the same home. If you need shared memory, use an external provider.
- Repeat it? Make it a skill. Skills don't eat the character budget.
- Check the files. If the agent says it saved something, open ~/.hermes/memories/ and look.
Common mistakes
- Expecting instant recall. The snapshot is frozen at session start. Something saved five minutes ago won't be in context until the next session.
- Stuffing memory with procedures. Long step-by-step instructions burn the character budget. Move them into skills.
- Replacing built-in memory. You can't, and you shouldn't want to. External providers run alongside it.
- Installing five providers at once. You won't know which one is helping. Add one, test, then decide.
- Trusting the agent's word. Check the memory files directly.
I run memory, skills and tools together inside my Agent OS so every agent keeps its context. You can see how it fits together in the Agent OS guide, and how different setups compare in Goldie Bench.
If you want the Hermes courses and the Agent OS itself, join the AI Profit Boardroom for $69/mo.
FAQ
What is the best Hermes Agent memory?
The best Hermes Agent memory is the built-in memory, MEMORY.md and USER.md, plus session search, with skills for anything you repeat. Add an external provider only when you hit a real limit.
Does Hermes Agent remember between sessions?
Yes. Built-in memory loads at the start of every session, and session search can pull up past conversations stored in ~/.hermes/state.db. Things saved mid-session appear from the next session onward.
How much can Hermes memory hold?
Per the Hermes docs, MEMORY.md holds about 2,200 characters (around 800 tokens) and USER.md about 1,375 characters (around 500 tokens). Past conversations live separately in session search, and external providers can add more.
Can two Hermes agents share memory?
Not through the built-in files. Memory is per profile, and the docs say not to point two agents at the same Hermes home. Use an external memory provider for shared memory.
Not sure which setup fits your workflow? Grab a free AI strategy session and we'll work out the best Hermes Agent memory setup for you.
