Best Ways to Create Portable Memory That Works With Any AI Assistant
7 min read · Updated Sep 5, 2026
There is no universal memory standard that every AI assistant reads from today, so nothing carries over automatically when you switch from ChatGPT to Claude to Gemini. What actually works right now is a plain-text or markdown memory file you write and maintain yourself, which you can paste or attach into any assistant, and MCP memory servers, a newer, growing option for tools that support the Model Context Protocol and want memory to persist and recall automatically without you retyping it each time.
Key takeaways
- No assistant reads another vendor's built-in memory. ChatGPT's memory, Claude's Projects, and Gemini's saved info are each locked to that one product.
- A plain markdown or text file you write yourself, your own facts, preferences, and project context, is the most portable memory format that exists, because every assistant can read plain text.
- MCP (Model Context Protocol), an open standard Anthropic introduced in November 2024, lets a growing number of assistants connect to the same external memory server, the closest thing to a real cross-assistant standard so far.
- Pasting a raw exported chat log from one assistant into another is not the same as portable memory, it dumps unfiltered noise (and sometimes another party's messages) into a fresh context with no curation.
- A memory file works well up to a few thousand words you curate by hand. Past that, a dedicated memory server with retrieval becomes worth the setup effort.
Why there is no universal memory format yet
Every major assistant vendor built its own memory feature as a product differentiator, not as a shared standard. OpenAI shipped ChatGPT Memory in February 2024 and expanded it through the rest of that year. Anthropic ships project-scoped context through Claude Projects. Google keeps saved personalization data inside its own account system. Each one stores what it remembers in a private, undocumented format, on that vendor's servers, readable only by that vendor's product. That is not an oversight, it is the natural result of memory being a retention feature for one company's app, and nobody has agreed on export formats because nobody who ships the memory feature benefits from making it easy to leave.
The option that actually works today: a memory file you own
A single markdown file you write and keep in a folder, or paste at the start of a new conversation, is the most portable memory format available, because every assistant can read plain text and none of them requires a special export format to accept it. This is exactly the pattern coding assistants like Claude Code, Cursor, and GitHub Copilot already use for project context files, a plain file at the root of a project that gets read automatically. Nothing stops you from doing the same thing for personal or professional context, manually, across any assistant that lets you attach a file or paste text.
- Stable facts: your role, the tools and stack you use, constraints that rarely change.
- Standing preferences: how you want answers formatted, what to avoid, tone you actually want.
- Active project context: what you are working on right now, kept current, deleted once the project ends.
- Decisions and their reasons: not just what you decided but why, so a new assistant does not re-litigate a settled question.
I think a memory file you write by hand beats every automatic memory-as-a-service product for most individual use. It is transparent, you can read every word of what an assistant "knows" about you, it is editable in seconds, and it works identically whether you are on ChatGPT, Claude, Gemini, or a local model over Ollama. The moment it stops working is when you need automatic recall across dozens of separate conversations without manually curating anything, at that point a hand-maintained file becomes a chore and a dedicated memory server starts earning its complexity. For a solo developer or a small team, that crossover point is further out than most people assume.
MCP memory servers: the closest thing to a real standard
The Model Context Protocol (MCP), an open specification for connecting an AI assistant to external tools and data sources, was introduced by Anthropic in November 2024 specifically to stop every vendor from reinventing its own private integration format. A memory server built on MCP stores facts outside any single vendor's walls and exposes them over the same protocol, so any MCP-compatible assistant can query it. By 2025 other vendors, including OpenAI's Agents SDK and Google's Gemini tooling, had added their own MCP support, which is what makes an MCP memory server meaningfully more portable than a vendor's built-in memory: the storage and the assistant are no longer the same company's product.
What built-in memory features get you, and the risk in moving raw exports
A vendor's built-in memory feature is genuinely useful inside that product: it recalls things automatically, with no file to maintain, and it is well integrated with that assistant's interface. The tradeoff is that it does not travel. Exporting a full chat history and pasting it into a different assistant is not portable memory, it is a wall of largely irrelevant back-and-forth, and if the export contains anything from a shared or team conversation, you are also handing another vendor's model text that was not written with that in mind.
The morning I lost a project mid-switch
In the fall of 2024, mid-way through building an extraction pipeline for a client, I switched from one assistant to another for a single feature that the first one kept getting wrong. I had no memory file, three weeks of architecture decisions lived only in one chat history, and I spent most of a morning, close to two hours, retyping context the first assistant already "knew": field names, why we rejected two earlier schema versions, which endpoints were flaky. Since then I keep a running markdown file for every active project, updated in a couple of minutes at the end of each session, and switching assistants mid-project costs me nothing more than pasting one file.
Comparing the three approaches
| Approach | Portability | Setup effort | Best for |
|---|---|---|---|
| Hand-maintained markdown memory file | Works with any assistant that reads text | Low, a few minutes to start and update | Individuals and small teams, most day-to-day use |
| MCP memory server | Works with any MCP-compatible assistant, configured per tool | Moderate, requires running or hosting a server | Automatic recall across many conversations without manual curation |
| Vendor built-in memory (ChatGPT, Claude Projects, Gemini) | Locked to that one product | None, works out of the box | Staying inside a single assistant long-term |
# Project memory: extraction-pipeline-v2
## Stable facts
- Stack: Python, Postgres, self-hosted n8n
- Source data arrives as nested JSON, not CSV
## Standing preferences
- Keep answers short, code first, explanation after
- Flag any breaking change explicitly
## Active context
- Currently debugging: duplicate rows after a retry
- Last decision: added an idempotency key, see below
## Decisions and why
- Rejected schema v1 (flat table): could not represent nested line items
- Rejected schema v2 (JSONB blob): unqueryable without a migration“The assistant you are talking to next week will not remember this conversation unless you make it easy for it to. That is not a limitation to work around once. It is a habit to build.”
Frequently asked questions
Frequently asked questions
Is there a way to create portable memory that works with any AI assistant?
Yes, in the sense that a plain text or markdown file you write and paste or attach works with any assistant that accepts text or file input, which is effectively all of them. There is no automatic, built-in equivalent that syncs across vendors on its own.
What is MCP and does it solve this?
MCP (Model Context Protocol) is an open standard, introduced by Anthropic in November 2024, for connecting an assistant to external tools and data, including memory servers. It solves part of the problem: any MCP-compatible assistant can query the same memory server. It does not mean every assistant does this automatically without configuration.
Can I just export my ChatGPT history and import it into Claude?
You can paste it in, but a raw export is an unfiltered transcript, not curated memory. Summarize it into the facts and decisions that matter first, otherwise you are spending context window on small talk and dead ends instead of the parts worth keeping.
Do I need a memory server, or is a text file enough?
A text file is enough for most individual and small-team use, up to roughly a few thousand words you are willing to curate by hand. Past that, or once you need recall across dozens of conversations without manual updates, a dedicated memory server becomes worth the setup effort.
Is it safe to put sensitive information in a portable memory file?
Treat it like any other document you would hand to a third party: no credentials, no API keys, no data you would not want leaving your machine, since the file's entire purpose is to be pasted into other people's AI products.