I use five AI assistants in a normal week: Claude, ChatGPT, and three Gemini accounts split across two workspaces and a personal login. Each one has a memory feature and each one remembers a different version of me.

One afternoon I asked Claude to tell me what I actually use it for, by category, so I could see where it kept starting from zero. The answer was uncomfortable. It had a decent picture of one project, a thin picture of another, and no idea what I had worked through in ChatGPT the day before. The other four assistants were in the same position, each holding its own slice and none holding the whole.

‍

Problem

A positioning framing I had refined in one assistant was still the old version in another, and that assistant used it with total confidence. A project status was accurate in one place and a few weeks stale in two others. Every new conversation opened with me re-explaining who was involved, what stage things were at, and which version of the plan was current.

The re-explaining was a visible cost but, the  real cost was that I could not trust any single assistant's idea of "current," because current depended on which app I happened to open last. Assistant memory is a silo by design. That works fine if you use one assistant, but I was seeing it breaks with two — not to mention five.

‍

Steps to Solve

• Pick one store every assistant can reach. I created a folder in Fulcra Files called /shared-memory/. Plain markdown, no database, nothing that any of the five apps could not read through Fulcra's MCP server.
• Write the read contract before writing any content. Every assistant opens index.md first, then at most three more files it points to. The index is a table: file, one-line description, last updated, status. That single rule stopped assistants from reading the whole folder, or worse, guessing.
• Separate current truth from history. topics/ holds one curated file per subject. sessions/ is append-only: one summary per conversation, with five fixed headings, including a Corrections section for anything the session discovered was stale.
• Make capture safe from anywhere. A session writes exactly one new file and edits nothing else. Two assistants checkpointing at the same moment cannot collide, because they never touch the same file.
• Wire each assistant its own way. Claude and ChatGPT connect to Fulcra directly and can read and write on their own. Gemini reads a mirrored copy of the index and hands me a summary to file. The store does not care which assistant wrote a file, and neither do the assistants reading it.
• Add a librarian. A weekly scheduled job merges new session files into the topic files, applies their Corrections, rebuilds index.md, and logs every action to an audit file. No assistant edits topics or the index directly. Only the librarian does.
• Let Claude build the tooling. I do not write code. Everything beyond the folder layout, the two Claude skills that capture sessions and file content from other assistants, the bundle initializer, the validator that checks every file against the schema, and the capture prompts for ChatGPT and Gemini, was built in Claude Code over a couple of sessions. I described the rules; Claude wrote and tested the implementation against my live Fulcra account.

‍

Solution

Now I have one folder in Fulcra with three parts: an index that is the first thing any assistant reads, topic files that are the current truth, and an append-only history of sessions that feeds them. Assistants write history. A librarian writes truth. The index says what exists and where to look, instead of repeating it.

I can open a ChatGPT conversation about a project and it reads the same status file Claude updated last week. I can ask Gemini about positioning and it lands on the framing that is actually current, not the one from three months ago. When any session finds something stale, the correction is made once, in one place, instead of being rediscovered in every app.

‍

Try It Yourself

You do not need the librarian or the schema on day one. Start with three things:

1. A Fulcra account. It is free to start with 5 GB, and Claude, ChatGPT, and Gemini all connect through the MCP server.
2. One index.md in Fulcra Files listing your active projects, one line each: what it is, what stage it is at, where the details live.
3. A one-sentence instruction in each assistant's preferences: "At the start of any substantive conversation, read /shared-memory/index.md in Fulcra first."

That is enough to stop the drift. Everything else in this post is what you add once you notice you want it.

If you want the full setup instead of building it up yourself, it is open source. The repo at github.com/keng009/fulcra-shared-memory (MIT) includes the two Claude skills (memory-checkpoint and memory-file-this), the bundle initializer, the validator, capture prompts for ChatGPT and Gemini, and a quickstart that takes about ten minutes with a Fulcra account and the Fulcra connector enabled in Claude.