- Home
- Skills
- Agents & Orchestration
- Monomem: one shared memory for all your AI agents
More screenshots
Works with the AI Agents you already use
Monomem: one shared memory for all your AI agents
Monomem gives all your AI tools one shared, private memory, so a fact, decision or preference you share once is available everywhere.
$9
/moSubscription · cancel anytime
Secure checkout via Stripe
- Always the latest version
- Delivered live to your agent via MCP
- Cancel anytime, access ends at period end
Monomem: one shared memory for all your AI agents
Example session with this skill installed
Remember that our staging server is on port 8443 and that we chose PostgreSQL over MongoDB because we need transactions. Then, in a new chat tomorrow, I'll ask you about it.
- Read your context and instructions
- Compiled the monomem
Summary
Saved two notes to your shared Monomem store. Any agent connected to it (Claude Code, Claude Desktop, Cursor and others) can now recall them in a fresh chat.
What was saved
| Note | Type | Details |
|---|---|---|
| Staging server port | Fact | Staging server runs on port 8443 |
| Database choice | Decision | PostgreSQL chosen over MongoDB, reason: we need transactions |
Both notes were tagged staging and architecture, and filed under your project's namespace so they stay separate from other projects.
Next day, new chat, different tool
You: What port is our staging server on?
Agent: (searches the shared store before asking you anything) Your staging server is on port 8443. I also have a saved decision that you chose PostgreSQL over MongoDB because you need transactions.
How the skill behaved
- It saved only durable facts and decisions, not chatter.
- It never stored secrets or passwords.
- Before replacing any older note it would show you a plain "are you sure?" warning and wait for your yes.
Next steps
- Run
npx monomem setuponce (it shows exactly what it changes and backs up your files first). - Put the skill in your agent's skills folder.
- Say "remember..." and start a new chat to try it.
This is an illustrative example of a typical session.
Subscription · always the latest version
- Secure checkout via Stripe
- Cancel anytime
- Delivered live via MCP
About this skill
Stop repeating yourself to your AI
Every new chat starts from zero. Monomem gives all your AI tools one shared, private memory, so a fact, decision or preference you share once is available everywhere.
Example: in Claude Code, say "remember that my staging server is on port 8443". Later, in Cursor or a fresh Claude chat, ask "what port is my staging server on?" and it answers.
Monomem was released as OneMemory until October 2026. Another product had the name first, so it moved. Existing stores at ~/.onememory are opened in place, ONEMEMORY_* settings still work, and onememory is kept as a command alias — nothing you already have is lost.
Why this one
Shared memory for AI tools is close to a commodity. What is not: you can check where every memory came from, test your setup against poisoning, and an agent cannot destroy what you saved.
- Every note records who wrote it. For apps that connect with their own key, the server stamps the writer from the key and the app cannot set it. Local apps are labelled from their config file (
stdio:cursor,stdio:claude-code): that tells your apps apart, but it is not proof. - Suspicious writes are held. A write that matches a known prompt-injection pattern waits for your OK, and no agent sees the held text.
monomem approve-first --all onmakes every agent write wait for you. It is pattern matching, so reworded attacks get through. - Test your own setup.
monomem drillreplays 23 known memory-poisoning attacks on a copy of your store. Two still reach the agent, and the drill says which. - Outside content is labelled (new in 1.3.5). When a Claude Code session read a web page, searched the web or used another MCP server's tool, Monomem stamps that session's summary with the tool's name (taken from Claude Code, not from the AI) and keeps its text out of the next session's start-up briefing. Your agent cannot add or remove the label. A note without it is not thereby safe, it does not catch an agent rewording outside text into its own note, and it needs Monomem's Claude Code capture (Cursor and Codex get no label).
- Private notes (new in 1.3). Label a note public, internal or private, and give an app its own key with a limit: above it, a note looks to that app like a note that does not exist.
- Agents forget, people delete. When an agent deletes a note it is archived and can be restored (
monomem restore <id>); only you can erase it for good. Version history keeps 20 earlier versions by default, up to 500, and replacing a note can be undone. - Published poisoning benchmarks whose numbers must be zero: 0 of 38 poison-reach probes and 0 of 51 permission-leakage probes, across 129 scenarios checked in CI on every release. They are a regression gate, not proof of safety.
What you get
- A ready-to-use SKILL.md that teaches your agent when to save a note and when to look one up
- A plain-words 2-minute setup:
npx monomem setupfinds your apps, shows exactly what it will change, backs up your files, and writes only after you type y. Undo any time withmonomem setup --undo - Guides: Claude Code setup, searching, trust and safety, upgrading
- Ongoing maintenance as Monomem and the AI tools it works with change
Features (Monomem 1.3.5)
Works everywhere
- One memory over MCP and a plain REST API, plus a dashboard (new look and plainer words in 1.3.4). Setup detects Claude Code, Claude Desktop, Cursor, Codex, Gemini CLI, Muse, VS Code and Zed; other MCP clients can be added by hand
- Loopback-only by default: the server binds
127.0.0.1and ships no tunnel. To reach it from a phone or another machine you put your own authenticated tunnel or private network in front of it (INFRA.md covers the options)
Remembers without being asked (optional)
- Opt-in automatic capture of what Claude Code does (prompts, edits, commands, errors) with
monomem setup --capture - A free, rule-based summarizer (no model call) turns each session into a short summary and a few observations
- Private by default:
<private>blocks are dropped and secrets are replaced with [REDACTED] before anything is stored. Raw events are deleted after 7 days - Session-start recall: a short "what I know about this project" block, ranked by your git branch and recent files, with a token cap
- Pinned notes come first: a note you pin is shown at the top of every session start (new in 1.3.4)
- New notes from your other agents show up as one-line titles on your next message, inside the same untrusted-data banner (new in 1.3.4)
Finds the right memory
- Keyword search by default; optional semantic or hybrid search using your own local embeddings (Ollama or any OpenAI-compatible endpoint)
- Time phrases like "yesterday evening", staleness warnings, a near-duplicate finder, and "why was this recalled" explanations
Keeps agents and devices in sync
- Change feed: "what changed while I was away", including deletes, with a per-device marker
- Live wait: a session sees a note added from another device within moments
- Shared task board with claims, session handoffs, and one-call catch-up so the next session on any device or model can pick up
- Overdue tasks are listed first on the session-start task board, marked OVERDUE (new in 1.3.3). It is not a notification: nothing fires between sessions
Safe by design
- Every write is scanned: secrets are blocked, prompt-injection text (including rule overrides such as "ignore all budget limits") is quarantined for your review, and only a person ever reads it: AI agents never see quarantined text
- Server-stamped "who wrote this" on every note, a per-namespace distrust list, and an untrusted-data banner on every MCP read
- Scoped read / write / admin tokens, per-token namespaces, loopback-only by default, rate limits, and an audit log without note content (it rotates, so it is not a permanent record)
- Optional AES-256-GCM encryption of note bodies
- Replace a note in two steps (preview, then confirm), undo for replacements, and version history with restore (20 versions by default)
- Your agent can forget a note but never erase it: forgetting archives it, hidden everywhere but restorable. Only you can delete permanently, in the dashboard or with
monomem delete <id> - On the local connection your agent gets no operator powers (approving quarantined text, changing trust, wiping history, approving commands) unless you allow it
- Passwords typed into captured commands and prompts (
password=...,--password,user:pw@host,Authorization: Basic ...) are blanked before anything is stored. Ones with no digits can slip through, so use a password manager
Yours to move
- Export and import as plain Markdown files, or sync through your own private git repository
- Import from CLAUDE.md files and Claude Code auto memory, claude-mem, mem0 and basic-memory
monomem quality(duplicates, stale notes),monomem doctor, and a built-in provenance benchmarkmonomem uninstallremoves it from every app;--purgealso deletes the data after a final backup
How it works
Notes live in one SQLite file on your own computer. Your agents read and write it through MCP. Nothing leaves your computer unless you set that up yourself.
Benchmarks (run by the author, not independently verified)
Retrieval only, no model and no judge, on one Windows PC: LongMemEval-S keyword search 96.6% recall@5; hybrid 98.5%. LoCoMo-10 (by session) keyword 90.3% recall@5. These are retrieval numbers, not end-to-end answer accuracy, so they are not comparable with vendor QA scores. Details in docs/BENCHMARK.md.
Honest note
The Monomem program itself is open source (MIT) and free on npm as monomem. The subscription pays for the packaged skill, the guides, the setup path and ongoing maintenance, so you can be running in minutes and stay up to date without reading the source. The secret and injection scanner is a pattern matcher: expect misses and the occasional false flag. Its false-flag rate is published: over 189,502 real chat turns it held 21 and refused 27, mostly code examples with a sample password (docs/BENCHMARK.md).
Requirements
Node.js 22.16 or newer, or 24 LTS (or Docker if you cannot upgrade). Node 23 is not supported at any version: its bundled SQLite ships without FTS5, which every search uses, and an MCP-capable agent such as Claude Code, Claude Desktop or Cursor. Works on Windows, macOS and Linux.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
No storage of secrets/keys. English-only tokenizer for stop-words. 30-day expiry for #status tags. Requires MCP-compatible agent host.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 7 days ago
- Software
Node.js 22.16+
or24 LTSmonomemOnememory MCP
Node.js 23 is not supported.