# Can Claude and Meta's Muse share one memory? Claude and Meta's Muse can share one memory: both now add the same store as a connector, and Muse builds its own when asked. Four options, and how to choose. Source: https://usewire.io/answers/share-memory-between-claude-and-muse/ Last reviewed: 2026-09-24 --- **Short answer.** Yes, by connecting both to the same store. Claude adds it as an MCP connector. Muse, as of September 2026, builds a custom connector when you ask and give it the address, then takes a key through a secure prompt. A Wire container works with both. Keep them separate when nothing you tell one is worth the other knowing. Yes, by connecting both to the same store. Muse remembers you rather than your work: it holds preferences, habits, and personal facts, and almost nothing about the decisions and documents you spend your day on. Claude holds the opposite. As of September 2026 both reach an outside store the same way, as a connector: Claude over MCP, and Muse by building a custom connector when you ask it to, so one store such as a [Wire container](https://usewire.io/solutions/portability/) can serve both. ## Why this happens Every assistant has its own memory, and none of them sync with each other. The separation is a side effect of how they were built rather than a deliberate strategy. Each memory is written in whatever format that assistant's own retrieval expects, and nothing in that format can be addressed from outside. Export usually exists, so no vendor can be accused of locking you in. None of them offer sync. As soon as you export, you have two copies. They drift apart from then on, and nothing reports it, because neither store knows the other exists. [Context portability](https://usewire.io/glossary/context-portability/) is the name for what is missing, and it is why "just export it" answers the question for one afternoon and not for the second week. A second reason is specific to this pair: the two products remember different categories of thing. A personal agent optimizes for the details you mention in passing, because those make it feel like it knows you. A work agent is handed material deliberately: a repository, a spec, a thread. Even a working connection between the two stores would be moving two different kinds of record, so the question to answer is which facts are worth having in both places rather than how to merge the two memories. ## What Muse remembers, and where it keeps it Muse's memory is files, and you can read them. Meta gives each user a dedicated Linux virtual machine in the cloud, and its [security write-up for Muse](https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse) says that "you can inspect, edit and download these files freely, including Muse's memory about you". Meta's [privacy help page](https://www.meta.com/help/artificial-intelligence/1047255454427887/) adds that you can see what Muse remembers by asking it or by opening files like `MEMORY.md`. That is unusually open for a consumer assistant, and it is what makes the manual-export option below possible. Meta's [launch announcement](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) says Muse "remembers what matters to a person" and acts on details mentioned only once, and that you can tell it to forget specific things it has learned. A Confidential VM that cryptographically locks Meta itself out of that data is announced for later in 2026, not shipped, so at the moment the isolation is a policy and an architecture rather than a cryptographic guarantee. Two things remain unconfirmed: whether Muse's memory merges with the Meta AI memory you may already have, and whether it can import memory exported from another assistant. Meta has said neither. ## How Muse connects to things outside Meta Muse connects to outside services through connectors, and you add one by asking for it. Meta's [connectors help page](https://www.meta.com/help/artificial-intelligence/1687253048996149/) says you can turn on a listed connector under Settings, then Connectors, or by asking ("Connect my Gmail"), and that for a service not on the list "you can ask Muse to create a Custom Connector", with the credentials kept in its Secure Credentials Store. Meta has also [opened connector submissions to developers](https://muse.ai/platform), so the list is growing. A store that speaks MCP goes through the custom connector. As of September 2026, you give Muse the server's address and say how it authenticates, and Muse builds the connector, tests it, and asks for the key through a credential prompt outside the chat. In our own use this month, Muse set the connector up itself when asked. Hands-on reports say custom connectors take a key rather than completing an OAuth sign-in, so plan on a key. Meta also says it "doesn't review custom connectors", which makes the scope of that key your main protection. The remaining constraints are about access. Muse launched in the US on September 8, 2026, on iOS, Android, muse.ai, and inside WhatsApp, and [reached Canada on September 18](https://www.iphoneincanada.ca/2026/09/18/metas-muse-ai-agent-is-now-available-in-canada/). A free tier sits alongside subscriptions [reported at $20 and $100 a month](https://www.cnbc.com/2026/09/08/meta-personal-ai-agents-public-reckoning-privacy-safety.html), and coverage of the launch notes that a payment card is required even on the free tier. ## Your options **One store both agents reach** is what a Wire container is, and it is the only option where something written by one agent is readable by the other without you in the middle. Claude and Muse both add the container as a connector, and Muse holds a key you scope to that one container and can revoke. It earns its place when the material genuinely overlaps. The risk is that Muse builds the connector itself, so a Muse update can change how it behaves. **Keeping them separate on purpose** is a real answer and often the right one. If Muse holds restaurant preferences and flight seats while Claude holds architecture decisions, a shared store adds a moving part and gains you nothing. It fails when you find the overlap later, after explaining the same fact to both agents more than once. **A shared document both are pointed at** costs nothing and works well while it is small. One page of standing facts, edited by Claude, read by Muse through a link. It breaks on growth and on concurrent writes. Past a few thousand words every question reads the whole file, and two agents editing one document have no way to merge their edits. **Exporting Muse's memory files by hand** is the option Meta's own design makes possible, and it is worth doing once. Download the files, read what Muse has decided about you, carry over what still matters, paste it into Claude. It fails as a habit on the second run, which is the general case covered in [moving your AI context between tools](https://usewire.io/blog/how-to-move-your-ai-context-between-tools/). ## How to decide **Does anything you tell one agent change what the other should do?** If the honest answer is no, stop here and keep them separate. If you can name three things in the last month, you have an overlap worth connecting. **Who writes, and how often?** A store that only Claude writes to and Muse reads is a much easier build than one both write to, and it covers most of the real cases. If both need to write, rule out the shared document immediately, because it has no way to handle two writers. **How much do you trust a connector Muse built itself?** Meta does not review custom connectors, and the feature is weeks old. If a week of a broken integration would be an annoyance, connect them with a key scoped to one container. If it would be a problem, keep the manual export until you have run the connection yourself for a while. Two answers point at Wire: you can name things you told one agent that the other should know, and a week of a broken integration would only be an annoyance. ## What to do next If you want it connected today, the sequence is short: create one store, add it to Claude over MCP, ask Muse to add the same address as a custom connector and enter a scoped key when asked, then write one entry from Claude and have Muse find it and write one back. If you are not ready for that, download your Muse memory files this week and read them. Ten minutes shows you which categories of thing Muse has been keeping, and it is the cheapest way to find out whether the overlap is real. Wire is our version of the shared store. It is one permissioned container that Claude and Muse both add as a connector, which is the [portability](https://usewire.io/solutions/portability/) case in its narrowest form. It is worth the setup when both agents genuinely work on the same material. When they do not, the second option on this page costs nothing and you should take it. Why any of this matters either way is in [why memory lock-in blocks your own product](https://usewire.io/blog/ai-memory-lock-in-blocks-your-own-product/): a memory you cannot address from outside cannot be moved. --- Sources: [Introducing Muse](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) · [Security and safety for AI agents: our approach with Muse](https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse) · [How Muse works with Connectors](https://www.meta.com/help/artificial-intelligence/1687253048996149/) · [How Muse handles your privacy, safety and security](https://www.meta.com/help/artificial-intelligence/1047255454427887/) · [Build a Muse connector](https://muse.ai/platform) · [Muse is now available in Canada](https://www.iphoneincanada.ca/2026/09/18/metas-muse-ai-agent-is-now-available-in-canada/) · [Meta pushes into personal AI agents](https://www.cnbc.com/2026/09/08/meta-personal-ai-agents-public-reckoning-privacy-safety.html) · [Meta debuts its Muse AI agent](https://techcrunch.com/2026/09/08/meta-debuts-its-muse-ai-agent-will-consumers-trust-it/) ## Options at a glance | Option | What it is | Best when | Breaks when | | --- | --- | --- | --- | | One store both agents reach (Wire) | A single permissioned container that Claude and Muse both add as a connector, so each reads and writes the same entries. | Both agents work on overlapping material, and you want a decision written by one to be readable by the other without you retyping it. | Muse builds its custom connectors itself and Meta does not review them, so re-test the connection after a Muse update. Muse is also available only in the US and Canada as of September 2026. | | Keep them separate on purpose | Muse keeps life context in its own memory, Claude keeps work context in its own, and nothing crosses. | The two halves genuinely do not overlap, or your employer's rules make a consumer agent the wrong place for work material. | The two halves overlap after all: travel that is also work travel, a contact who is both, a decision you explained once and now explain again. | | A shared document both are pointed at | One long Markdown or Google Doc that Claude edits and Muse reads through its browser or a link. | You have one page of standing facts that changes weekly, not daily, and you want zero new accounts. | Growth breaks it. Past a few thousand words, both agents read the whole file to answer anything, and neither can write into it safely without clobbering the other. | | Export Muse's memory files by hand | Download the memory files from Muse's VM, paste the parts worth keeping into Claude, repeat when it drifts. | A one-time move, or an audit of what Muse has actually learned about you. | The second export. After that there are two copies that disagree, and keeping them in step is your job. | ## FAQs **Does Muse support MCP servers?** Yes, through custom connectors, as of September 2026. You ask Muse to connect a service and give it the server's address, and Muse builds the connector, tests it, and keeps the key in its Secure Credentials Store. Meta's help pages describe custom connectors without naming MCP, but they work against remote MCP servers in hands-on use, ours included. **Can I import my Claude or ChatGPT memory into Muse?** Meta has not said. The documented direction is the other one: memory lives as files inside your Muse VM, including a MEMORY.md you can open, and you can inspect, edit, and download them. Whether Muse ingests another assistant's exported memory, or merges with the Meta AI memory you already had, is unconfirmed either way, so treat any claim in either direction as unverified. **Is it safe to give a consumer AI agent an API key to my own data?** It is reasonable with a key scoped to exactly what you want the agent to reach, with an expiry, and revocable in one action. Muse keeps the key in a credential store the model cannot read, but Meta does not review custom connectors, so the scope of the key is your real protection. A key that opens one container limits the damage to that container. **What does keeping two agents on one store actually cost?** Writing is free and reading is the part that costs. On Wire, writing an entry and deleting one are free, browsing costs 1 credit per call, a hybrid search costs 5, and storage is 1 credit per 1,000 entries per day. A personal store that two agents query a few dozen times a day sits inside the free starter grant for weeks. **Muse is not available in my country. Does this work elsewhere?** The shared store works anywhere, but Muse is available only in the US and Canada as of September 2026. Any agent that can add the same container as a connector reads and writes the same entries, so the pattern survives swapping Muse for whichever consumer agent is available where you are. Only the connection step changes.