Your AIs don't talk to each other. Here's how to give them one shared memory in 2 minutes

Pryzm Memory — Tue, 06 Oct 2026 05:10:43 +0000

Monday, 9:14 pm. I tell ChatGPT about a new client: Claire Morel, Atelier Lumen, showcase website, €4,800, delivery end of November. She wants a clean look. No purple.

Tuesday morning, I open Claude to write her proposal.

Claude has never heard of Claire.

So I paste the context again. And again on Wednesday, in another chat. And again in Cursor. If you use more than one AI assistant, you know this tax: every assistant has its own memory, and none of them share it. You are the copy-paste layer between your AIs.

This post fixes that in about two minutes, with no code and nothing to install.

The trick: a memory that lives outside your assistants

The Model Context Protocol (MCP) lets an assistant call external tools. Most people use it for GitHub or Notion. But a tool can also be a memory: a server with store, search and recall functions.

Host that server at a URL (a remote MCP server), connect every assistant to it, and something simple happens:

What Claude saves, ChatGPT can find. What ChatGPT learns, Cursor already knows.

One memory, all your AIs.

Every memory, grouped by client and project. (The app UI is in French for now: it's built in France.)

I'll use Pryzm Memory, the remote memory server I'm building. The steps are the same for any remote MCP server with OAuth.

Step 1: Claude (30 seconds)

claude.ai → Settings → Connectors
+ → Add custom connector
Name: Pryzm, URL: https://pryzm-memory.com/mcp → Add
Connect → sign in (OAuth)

Works on the free plan (one custom connector), web and desktop.

Step 2: ChatGPT (1 minute)

ChatGPT connects to remote MCP servers through Developer mode (beta, web, paid plans).

chatgpt.com → Settings → Apps
Advanced settings → turn on Developer mode
Create an app with https://pryzm-memory.com/mcp, authorize with OAuth
In a new chat: + → Developer mode → enable the app

Each assistant gets its own token. Revoke one, the others keep working.

Step 3: the part everyone skips

Connecting gives your AI access. It doesn't make it use the memory. Models are lazy: if you don't tell them to check, they won't.

Paste this in Claude (Settings → General → personal preferences) and ChatGPT (Settings → Personalization → Custom instructions):

Pryzm Memory is my main memory, shared across all my AIs.
1. At the start of every conversation, call memory_whats_up before answering.
2. Before any answer that depends on my context, preferences or projects, call memory_search.
3. Save right away with memory_store every decision, preference, goal or project update.
4. If memory contradicts your assumptions, memory wins.
5. If my message starts with ##, it is private: save nothing from it.

That last line is my favourite: start a message with ## and nothing from it is ever saved.

The moment it clicks

In ChatGPT:

Quote for Claire Morel (Atelier Lumen): showcase website, €4,800, delivery end of November. She wants a clean look, no purple.

ChatGPT calls memory_store. Now close it, open Claude, and type only:

Draft the proposal email for Claire.

Claude calls memory_search, finds the quote, and writes the email with the right price, the right deadline, and no purple. You never repeated a word.

The memory Claude found, saved the day before from ChatGPT (demo data).

That's the whole point. Not a smarter model. A model that remembers what you already told another one.

Under the hood (for the curious)

A memory is only as good as its search. Asking "what did Claire want for the design?" has to find a note that says "clean look, no purple", which shares almost no words with the question.

So the search runs two engines at once:

Keyword search (BM25) catches exact names, prices, dates.
Semantic search (pgvector embeddings) catches paraphrases, and questions asked in another language than the note.

Both result lists are merged with reciprocal rank fusion, then a reranker reorders the top results. All of it on PostgreSQL.

I publish the retrieval benchmarks in the public repo, including where it loses. On the public LoCoMo benchmark (1,532 questions, retrieval only), the evidence lands in the top 5 for 64% of questions, against 52% for keyword search alone and 37% for semantic search alone. The known weak spots, measured and published: outdated facts that still rank first, questions that need two memories at once, and questions with no answer in memory. I'd rather you read them from me than discover them yourself.

Honest limits

ChatGPT Developer mode is web only and not on the free plan.
The model still decides when to call the tools. The instructions above make it far more consistent, not perfect.
A remote memory means your notes live on a server. Pryzm is hosted in Germany and encrypted at rest. If your data must never leave your machine, use a local MCP server instead.

Works elsewhere too

Same URL in Claude Code, Cursor, VS Code (Copilot), Mistral Le Chat, Perplexity and most MCP clients with OAuth. Guides: https://pryzm-memory.com/en/connect/

The dashboard: what all your AIs know about your work, in one place (demo data).

Try the Claire test with your own AIs and tell me in the comments what breaks. I read every one.

DEV Community: Pryzm Memory