omnemo MCP server
Commit Nov 14, 2026Provides a shared memory layer for AI agents on Omarchy, allowing different harnesses to save and recall information across sessions.
Also known as GuyMannDude/omnemo
- Stars
- 1
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 14, 2026
- License
- MIT
What is omnemo MCP?
omnemo provides a shared memory layer for AI agents on Omarchy, allowing different harnesses to save and recall information across sessions.
Summary generated from the project's README
About omnemo MCP server
Providing a shared memory layer for AI agents on Omarchy allows different harnesses to save and recall information across different sessions. This system enables agents like Claude Code, Codex, OpenCode, Crush, and Gemini to access a unified memory store. Instead of each agent starting with a blank slate, they can access information previously stored by other tools on the same machine. This creates a persistent context where an agent can remember past interactions, user preferences, or even previous software crashes and their fixes.
Users interact with this memory through standard conversation with their AI helper. The server operates locally, storing data in the ~/.local/share/omnemo directory, and does not send information to external services. It uses a local embedding model for similarity-based recall, which mimics human memory by blending recency, importance, and similarity. The setup is integrated into the Omarchy ecosystem via skills and plugins, and can include an optional memory count widget in the Omarchy bar.
What you can do with it
- 1Cross-agent information sharingA developer can have Claude Code store a specific fact and then ask Codex to retrieve it later.
- 2Automated crash diagnosisAn agent can use past memory to identify why a program crashed and suggest a previously successful fix.
- 3Personalized AI interactionsAn AI helper can remember your name or project details to provide more relevant assistance in new chats.
- 4Managing sensitive dataThe system can be instructed to forget specific information upon request.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 16, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableguymanndude/omnemo is public and reachablegithub.com/guymanndude/omnemo ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-14github.com/guymanndude/omnemo ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubOmnemo — memory for Omarchy
Agents come and go. The memory abides.
One shared memory for every agent harness on the box.
Actively maintained, MIT licensed, and running on Omarchy 4.5. Install from
main with INSTALL.md. Want a friendlier desktop to go with it?
Omnemo OS sets up both.
Omarchy made coding agents system citizens. Omnemo gives them the one thing they still wake up without: memory. Claude Code remembers what Codex did yesterday. Your agent knows this app crashed before — and what fixed it.
Built just for Omarchy. Not a fork, not a patch — a
glove-fit layer made entirely from Omarchy's own extension points (skills,
user plugins, hooks), so it survives every omarchy update.
What you get
- Shared recall across harnesses — one memory store, spoken over MCP, registered with Claude Code, Codex, OpenCode, Crush and Gemini (Pi gets the skill; its MCP registration is not supported yet). Save in one, recall in any.
- Human-like memory, not a fixed database — recall blends similarity, recency, and importance; memories decay at different rates by kind, the way yours do.
- A memory skill dropped into
~/.agents/skills/and linked into the Claude Code, Codex and Pi skill dirs — those harnesses learn when to remember and when to recall, automatically. - Crashes with a past — Omarchy hands core dumps to your agent; Omnemo lets the agent answer "this crashed before, here's what fixed it."
Optional extra, off by default: a memory count in the Omarchy bar —
omnemo setup --bar (or omnemo bar on any time; omnemo bar off removes it).
Install
Fastest path: pick a default agent if you haven't (
omarchy default agent claude, orcodex, …), open it (omarchy agent) and say "Install Omnemo — follow INSTALL.md." It will do the rest.
Manual instructions: INSTALL.md.
Using your memory
There's nothing new to learn. Talk to your AI helper (Claude, Codex, OpenCode or another) the way you always do. When you tell it something worth keeping, like your name, how you like answers, or a project you're working on, it saves it and tells you in one sentence that it did. Next time, in a new chat or even in a different AI app on the same computer, it already knows.
- Want something gone? Just say so: "forget that I…". Your helper finds it, confirms which one, removes it, and tells you.
- Personal things (health, money, family): your helper asks before it saves them.
- Where it lives: in one folder on this computer,
~/.local/share/omnemo. Omnemo itself sends nothing anywhere. When your helper reads a memory to answer you, its AI service sees that memory, the same as anything you type to it. - Starting fresh: to wipe everything, copy that folder somewhere safe, then delete it.
Prove it
Memory is durable the instant a save returns — a crashing harness loses nothing. Don't take our word for it:
python3 examples/crash_demo.py
It opens a real MCP session against omnemo serve, saves a memory, SIGKILLs
the server the moment the save is acknowledged (no shutdown, no flush), then
recalls the memory from a fresh process.
Cross-harness recall, live
The core promise — save in one harness, recall in any — is a two-command demo on a box with two authenticated harnesses:
# Ask Claude Code (Anthropic) to remember something
claude -p 'Use the omnemo save tool to store this memory: the electronics
repair shop favorite song was "Everything'\''s Broken" by Bob Dylan' \
--allowedTools "mcp__omnemo__save"
# Ask Codex (OpenAI) — a different vendor's harness, fresh process
codex exec 'Using only the omnemo MCP recall tool: what was the favorite
song at the electronics repair shop?'
Run live on the reference Omarchy 4.0.1 box:
mcp: omnemo/recall started
mcp: omnemo/recall (completed)
codex
"Everything's Broken" by Bob Dylan.
One memory, written by one vendor's agent, recalled by its competitor. The memory belongs to the machine, not the harness.
What to expect on a low-spec box
Measured 2026-10-30 on a 2016 laptop: Intel i3-6100U (2 cores / 4 threads), 3.7 GB RAM, Omarchy 4.5.0rc2 (rc channel), Omnemo 0.2.1.dev3, timed over the real MCP stdio path.
| What | Number |
|---|---|
Cold start (spawn omnemo serve → first recall answered) |
2.6 s |
| Warm recall, 20 calls (4 memories) | p50 167 ms, p95 202 ms |
| Warm recall, 20 calls (1,000 memories) | p50 136 ms, p95 148 ms |
| Save | p50 79 ms, p95 139 ms |
| Server memory (RSS), warm | ~330 MB (337 MB at 1,000 memories) |
| Store on disk | 32 KB at 4 · 236 KB at 100 · 2.1 MB at 1,000 memories |
One-shot CLI omnemo recall (loads the model each time) |
1.6–1.9 s |
Most of the memory and the cold start is the local embedding model
(BAAI/bge-small-en-v1.5). Each harness runs its own server, so budget
~330 MB per open harness on a 4 GB machine. The cold start was measured with
the model already in the OS file cache; the first run after a reboot is slower.
Status
v0.2.0, proven on a clean Omarchy 4.0.1 box (September 2026); partly re-verified on a clean Omarchy 4.0.4 box (2026-10-09). Fresh install to green in minutes: MCP registered with all five supported harnesses (verified read-back, not just name-matched), skill linked, bar widget live.
Re-proven on 4.0.4: the manual INSTALL walk, crash-durable save
(examples/crash_demo.py), the skill surviving repeated omarchy update runs,
theme switches and shell restarts, the bar plugin on the 4.0.4 Quickshell bar,
and local-first operation with the network cut.
Proven on 4.0.1 only, not yet re-run on 4.0.4: the agent-driven install
("Install Omnemo — follow INSTALL.md"), save-in-one-harness / recall-in-another,
and survival of harness upgrades. The crash-diagnosis hook is a real 4.0.4
extension point; no agent was driven through a crash. OpenCode, Crush and
Gemini are registered but were not driven. See SPEC.md for the
design.
Lineage
Omnemo is the memory core of Mnemo Cortex — "memory that works more like a human than the others" — rebuilt clean and cut down to just memory, then fitted to Omarchy natively. MIT licensed, like Omarchy itself.
omnemo MCP questions
How does Omnemo handle memory across different AI tools?
It provides a single shared memory store that can be accessed by multiple agent harnesses like Claude Code, Codex, and Gemini via MCP.
Is my data sent to the cloud?
No, Omnemo is local-first and stores everything in a folder on your computer; it sends nothing externally.
Can I delete specific memories?
Yes, you can instruct your AI helper to forget specific information, and it will find and remove it.
Does Omnemo work if the agent crashes?
Yes, memory is durable immediately upon a successful save, meaning a crashing harness will not lose data.