Two AIs, two memories: what if you kept just one?
Stéphane Commenge, with Claude ·
The Fact
According to the 2026 Baromètre du numérique (Digital Barometer) from Arcep and Crédoc, 51% of generative AI users use at least two. Each one keeps its own memory, and portability between vendors remains, according to the Data Transfer Initiative, inconsistent and very incomplete.
The Signal
The danger is not having several AIs. It's leaving your memory in pieces at each one.
An AI is a very good pen. It is not a filing cabinet.
The Action
Go read what your AI has retained about you, in its settings. If it offers an export, do it and keep a copy at home.
It's your filing cabinet, until something better comes along.
Series "AI memory", episode 5. Previous episode: when AI still believes what is no longer true.
Claude for working. Perplexity for searching. ChatGPT for testing.
Three AIs, one single question: "where do we stand?"
When I start a search on Perplexity, it already knows where the project stands. It hasn't read Claude. They read the same memory. Like a colleague who walks into the meeting having read the minutes. A rare species.
And I'm not an isolated case.
One user in two has several AIs
The 2026 Baromètre du numérique (Digital Barometer) from Arcep (France's telecoms regulator) and Crédoc (a French research center on living conditions), published on February 9, 2026, is based on 4,145 people aged 12 and over. Among generative AI users, 51% use at least two. ChatGPT is the main tool for 63% of them, Gemini for 13%.
In the United States, a survey by Inflection AI of 1,000 Americans, published on July 21, 2026, puts the average at 4.2 assistants. A caveat: Inflection AI is an industry player. Cross-check this figure.
Two AIs, two memories. Each one keeps what you've entrusted to it. One has no idea what you told the other.
Moving from one to the other
On September 1, 2026, Chris Riley published "The shifting locus of AI portability" for the Data Transfer Initiative, an organization that works on data portability. His starting point: your personal data mostly lives in the agent's memory, not in the model. That's what you need to be able to take with you.
His verdict on current portability: "inconsistent and very incomplete". At the date of the post, Anthropic lets you import and export Claude's memory, Google publishes an import guide for Gemini, and OpenAI had announced nothing.
On Google's side, importing memory and histories into Gemini was introduced on March 26, 2026. It was not available at launch in the European Economic Area, the United Kingdom and Switzerland. For a French user, door closed at launch.
A common format? Not yet
On July 6, 2026, MacPaw Research proposed an open format called "Portable Memory". A preprint. A proposal, not an adopted standard.
Let's sum up. Several AIs per person. One-off imports at some vendors, not at others. A common format still on paper.
The danger is not having several AIs. It's leaving your memory in pieces at each one.
You don't keep your files at the pen maker's. You keep them in your filing cabinet, and you pick the pen that suits. An AI is a very good pen. It is not a filing cabinet.
How two AIs read the same memory
Through a connector. The MCP protocol (Model Context Protocol) lets an AI call an outside service: look up a document, read a calendar, consult a memory. When that service is online, we speak of a remote MCP connector. Every AI that accepts it can plug into it.
The memory no longer lives at the AI's vendor. It lives apart, and the AIs come to read it. An import copies a memory once. A connector shares it with every question.
A tip
Go read what your AI has retained about you. With many assistants, the memory can be viewed in the settings, and some offer an export. Keep a copy at home. It's your filing cabinet, until something better comes along.
If you switched AIs tomorrow, what would you lose? The thread of a project followed for months, the tone you taught your assistant, the decisions made along the way, the names of the people you deal with?
What I'm building with Mycelora
The filing cabinet. A single memory, yours, plugged into Claude, ChatGPT, Perplexity, Mistral Vibe and any AI that accepts a remote MCP connector.
Fully functional with the plugin on Claude and in ChatGPT's desktop app. Elsewhere, a memory you can consult.
Exportable. It follows you if you change tools.
Hosted in Paris at Scaleway.
The beta is open: mycelora.ai
What's next
The series continues, following the news on AI memory.
Sources
- Arcep and Crédoc, Baromètre du numérique 2026, February 9, 2026: arcep.fr
- Chris Riley, Data Transfer Initiative, "The shifting locus of AI portability", September 1, 2026: dtinit.org
- Google, March 26, 2026 post on importing into Gemini: blog.google
- MacPaw Research, "Portable Memory", July 6, 2026: research.macpaw.com
- Inflection AI, survey of 1,000 Americans, July 21, 2026 (industry vendor)