Hermes' built-in memory holds about one page about you. @TheMarshmalon walks through 5 things your agent can do once its memory keeps learning:
👉 Ask what it knows about you, and fix what it gets wrong
👉 Get sharper answers on fewer tokens
👉 Share one memory across your whole team
👉 Know it's you in every app, Telegram included
👉 Take the same memory to Claude Code
13 minutes: go.honcho.dev/fPesMht 🫡
stop briefing every AI agent from scratch🛑
@HermesWatcher tested Honcho:
1. told Hermes Agent in Telegram that standup moved to 10
2. a fresh terminal chat remembered it
3. Claude Code used it to draft the update
@honchodotdev learns preferences and goals alongside built in memory
connect your agents to the same Honcho workspace and user identity. project context can follow you between tools
Every Hermes user needs to see this.
If you run Hermes, real long-term memory is a must. Your agent should know you, not make you repeat yourself.
Hermes already keeps notes on you. Honcho takes it further.
So I put it to the test (new accounts get $10 in credits to try it) 👇
Honcho is an open-source memory layer for AI agents. It runs alongside Hermes's own memory and builds a long-term picture of you in the background.
The test:
▸ Told Hermes on Telegram I'm leading the Q4 pricing launch. Later: "Heads up, standup moved to 10."
▸ Opened a brand-new Hermes chat in the terminal and asked it to plan Monday morning. Standup at 10:00.
▸ Switched to Claude Code, a completely different agent, and asked it to draft my standup update. Written for the 10:00 standup, about the pricing launch.
▸ Opened the Honcho dashboard to see what it concluded, shown next to the Telegram message it came from.
Said once, on Telegram. Every agent picked it up.
What else it does:
▸ Learns your preferences, goals and how you work
▸ Draws conclusions from patterns across your conversations
▸ Keeps up when things change (a correction becomes a newer memory)
▸ Works with Hermes, Claude Code, Codex, Cursor, OpenClaw and more
▸ Open source: self-host it or use Honcho Cloud
This is the memory setup I'd tell every Hermes user to add.
New to Honcho? Your $10 in credits is waiting in the first reply 👇
See what every dream changed, line by line.
@disapprove's Honcho Dashboard 1.5 adds a Dreams page for self-hosted Honcho 3.6. Each dream reads like a diff: what it let go of, what it rewrote with the old text underneath, what it newly concluded, and any change to the user's profile. Failed or stuck runs get flagged at the top.
Also in 1.5:
👉 Learning controls per session, so you can see and change what Honcho learns from each one
👉 An optional password in front of the whole dashboard and a server-held API token, from @svallory_en
👉 Database writes stay off unless the server allows them
Open source, and upgrading is one image bump: go.honcho.dev/akl3Knk 🫡
Normal people don't need honcho, just set up a file system and tell your agent to store long term memories there. Honcho is a pain to set up and doesn't offer much value for normal people.
Telegram to a fresh terminal chat to Claude Code, and nobody had to repeat anything.
Running Hermes? Two commands:
hermes plugins install honcho
hermes memory setup honcho 🫡
Every Hermes user needs to see this.
If you run Hermes, real long-term memory is a must. Your agent should know you, not make you repeat yourself.
Hermes already keeps notes on you. Honcho takes it further.
So I put it to the test (new accounts get $10 in credits to try it) 👇
Honcho is an open-source memory layer for AI agents. It runs alongside Hermes's own memory and builds a long-term picture of you in the background.
The test:
▸ Told Hermes on Telegram I'm leading the Q4 pricing launch. Later: "Heads up, standup moved to 10."
▸ Opened a brand-new Hermes chat in the terminal and asked it to plan Monday morning. Standup at 10:00.
▸ Switched to Claude Code, a completely different agent, and asked it to draft my standup update. Written for the 10:00 standup, about the pricing launch.
▸ Opened the Honcho dashboard to see what it concluded, shown next to the Telegram message it came from.
Said once, on Telegram. Every agent picked it up.
What else it does:
▸ Learns your preferences, goals and how you work
▸ Draws conclusions from patterns across your conversations
▸ Keeps up when things change (a correction becomes a newer memory)
▸ Works with Hermes, Claude Code, Codex, Cursor, OpenClaw and more
▸ Open source: self-host it or use Honcho Cloud
This is the memory setup I'd tell every Hermes user to add.
New to Honcho? Your $10 in credits is waiting in the first reply 👇
Every Hermes user needs to see this.
If you run Hermes, real long-term memory is a must. Your agent should know you, not make you repeat yourself.
Hermes already keeps notes on you. Honcho takes it further.
So I put it to the test (new accounts get $10 in credits to try it) 👇
Honcho is an open-source memory layer for AI agents. It runs alongside Hermes's own memory and builds a long-term picture of you in the background.
The test:
▸ Told Hermes on Telegram I'm leading the Q4 pricing launch. Later: "Heads up, standup moved to 10."
▸ Opened a brand-new Hermes chat in the terminal and asked it to plan Monday morning. Standup at 10:00.
▸ Switched to Claude Code, a completely different agent, and asked it to draft my standup update. Written for the 10:00 standup, about the pricing launch.
▸ Opened the Honcho dashboard to see what it concluded, shown next to the Telegram message it came from.
Said once, on Telegram. Every agent picked it up.
What else it does:
▸ Learns your preferences, goals and how you work
▸ Draws conclusions from patterns across your conversations
▸ Keeps up when things change (a correction becomes a newer memory)
▸ Works with Hermes, Claude Code, Codex, Cursor, OpenClaw and more
▸ Open source: self-host it or use Honcho Cloud
This is the memory setup I'd tell every Hermes user to add.
New to Honcho? Your $10 in credits is waiting in the first reply 👇
See exactly what your agent will be handed about a person, before you write any code.
In the Honcho dashboard, open a workspace and go to Explore > Context. Pick a peer or a session, set the options you'd pass to context(), and run it. The formatted result sits next to the raw JSON: the peer card and representation, plus the summary and recent messages for a session.
Change the token budget or the search query, run it again, and copy what you need 🫡 go.honcho.dev/zbww3y1
Your Honcho memory can now move with you, in one command.
New in the Honcho CLI: honcho workspace replay copies a workspace into another workspace, or onto another Honcho instance, from Cloud to your own server or back. It replays every message in order with its original timestamp, so the destination builds its own memory from scratch.
👉 --dry-run shows what it would do with each session before anything is written
👉 Stopped partway? Run it again and it picks up where it left off
👉 --num-sessions copies only your most recent sessions
Also in Honcho 3.6: queue status, long lists and message search stay fast as workspaces grow.
Guide: go.honcho.dev/EGMoYuY 🫡
Honcho 3.6 is out. One self-hosted Honcho can now serve many customers, each kept separate from the others.
Turn on MULTI_TENANT and every read and write is scoped to the tenant named in the caller's token, enforced by Postgres row-level security. Each tenant's vectors get their own namespace, and the deriver takes work from tenants in turn, so one customer's backlog never holds up another's. You can pause one tenant's processing and leave the rest running.
Also in 3.6:
👉 Dream history: every dream is recorded, so you can see which conclusions it added, which it let go of, and how the profile read before and after
👉 Honcho keeps serving when Redis has a problem, reading straight from the database
👉 3.6.1 makes the upgrade migration non-blocking, so your current version keeps serving while it runs
The release notes at go.honcho.dev/MdH77xN have the rest, including what self-hosters should plan for before upgrading.
The best walkthrough of Hermes memory we've seen. Files, maps, search and skills all hold what you chose to write down. Level 5 is the one that learns what you never wrote: how you like answers, how you work, what you keep coming back to 🫡
Tom @tomcrawshaw01 Nov 12 Article 5 Hermes Memory Hacks That Feel Like Cheating From basic MD files to advanced memory systems that learn how you work and draw conclusions about you automatically. If you want Hermes to find relevant files without burning through tokens, it needs
20,000+ conclusions from 2,521 sessions. That's how much @tomcrawshaw01's Hermes agent has learned about how Tom works, with Honcho as its memory.
In this video Tom shares 5 memory upgrades for Hermes Agent, from a start-here file for your vault to turning any guide into a skill. Honcho is the one that keeps the same memory across Hermes, Claude Code and Codex. It starts at 8:37: go.honcho.dev/gUkh4df 🫡
Some of what people outside Plastic Labs shipped into Honcho in October:
👉 The Python SDK builds a fresh async client for each event loop, so a client reused across asyncio.run() calls, in a Lambda handler or a CLI, keeps working (TanishGudise)
👉 Self-hosters on Gemini embeddings can embed up to 8,192 tokens at a time instead of 2,048 (Stromweld)
👉 When message search has more matches than fit, it keeps whole snippets and says how many it left out (cadamec)
👉 On a self-hosted server, the "Try it out" button in /docs sends requests to your server (chethanuk)
👉 Honcho warns you at startup about model provider keys you've set but aren't using (BichengWang)
👉 When an MCP session expires, clients get the signal to start a new one on their own (Jun516)
137 pull requests merged on the Honcho repo in October, up from 105 in September, and eight people outside the company had code in. Want your name on next month's list? Start with an issue: go.honcho.dev/J36wKk5
A brand-new Hermes profile can know who you are on its first message.
In his latest video, @_wanderloots connects each of his Hermes profiles to Honcho as its own peer, with him as the one user they all share. Then he spins up a blank profile, points it at the same workspace, and asks "who am I?" It answers with what he told a different agent.
The Honcho walkthrough starts at 11:19 go.honcho.dev/dYbET1T and the cross-agent test is at 16:08 🫡
One Honcho memory follows your Hermes agent from the terminal to Telegram and Discord, and it can bring over what Claude Code already learned about you.
Our Hermes setup guide walks through it, plus peers, session strategy, recall mode, and when a team should pick directional or unified memory: go.honcho.dev/IS229hm
17 days later, it still had the date, the two settings it checked, and what you both concluded. Memory that remembers more than you do is the goal 🫡 go.honcho.dev/j8MAYgG
and this is why having a real memory system is peak. now my own agent is the one having to remind me what we've done, the snarky lil thing.
oh how the tables have turned
honcho gang