Memory is table stakes in 2026. The interesting question is no longer whether your AI remembers you. It is which kind of remembering each tool is built for.
By Travis Sawyer, Founder · Published May 15, 2026
The word "memory" gets thrown around as if every product means the same thing. They do not. Here is the honest read on each.
Memory is on by default for personal tiers. The system extracts facts and preferences from conversations and surfaces them in later sessions. Memory is scoped within ChatGPT and is opt-out for training. Export is supported via the data-export flow. See OpenAI's memory FAQ for the current state.
Memory rolled out to all users in March 2026. The architecture is transparent: Claude uses file-based memory hierarchies and surfaces when it is using a memory rather than silently injecting it. Projects keep memory siloed by workspace. No memory on the Claude API or Claude Code by default. No consumer training on memory data by default. See the Claude 2026 memory guide.
Memory has been in the Gemini app since 2024 and is integrated with Google account context where users opt in. The newer Gemini Spark agent leak from May 2026 points to a more autonomous direction, but the consumer memory feature today is account-bound and visible to the user.
Persistent memory has been live since April 2025. The implementation uses vector embeddings, with user-visible memory controls. Not currently available to EU or UK users. xAI's memory feature is conventional in shape but well-controlled at the user level.
A different use case. Spaces are research workspaces where you collect documents and then ask questions against them. Recall is excellent, but the product is organizing research output, not maintaining a long-running relationship. Over 5 million Spaces have been created. It is the right tool for "remember everything about this project" and a different tool from the relational continuity question this piece is really about.
A separate category. These products are built around roleplay, simulated companionship, and emotional connection as the primary feature. Memory engineering is often genuinely sophisticated, but the use case is different from a general-purpose assistant. The companion category sits inside an active legal and regulatory conversation that is worth taking seriously on its own terms; I am not the right person to adjudicate it, and I am not going to pretend ReGild is in that category. We are not.
In March 2026, the three biggest assistants shipped memory import and export within weeks of each other. OpenAI added export. Anthropic added import from ChatGPT and Gemini. Google added the same on its side. Tom's Guide ran a clean walkthrough of the 60-second ChatGPT-to-Claude path.
The honest read on what actually transfers: facts, preferences, and instructions. Your name. Where you live. What you do. Your stated response-style preferences. What you have told the system you like.
What does not transfer: conversation history, behavioral patterns, the implicit understanding the model built about you over time, the voice that emerged from your prompting. Anthropic labels the import feature experimental, with the explicit caveat that imported memories may not always incorporate cleanly. That is honest of them and worth repeating: import moves what you said, not how the model responded.
This is not a complaint about the feature. Memory import is a real improvement. It does the half of the job that can be done by moving structured facts. It does not do the half that requires moving relationship context. That second half is what persona portability addresses, and it is a different architectural problem.
ReGild is not better at recall than the memory features inside ChatGPT, Claude, or Gemini. Their memory is well-engineered for their product. ReGild's category is different.
Every major tool builds memory inside its own walls. ChatGPT's memory is for using ChatGPT. Claude's memory is for using Claude. Gemini's memory is for using Gemini. Memory import moves facts between them; it does not move the voice.
ReGild is built around persona portability across providers. You build a persona inside ReGild, accumulate context, then run that persona on whichever model is best for the moment, through your own API keys. The persona's voice and accumulated context travel with you when you switch the underlying model. The architecture is documented on our architecture page.
This is a separate question from which provider has the best memory feature. It is the question of where the relationship lives. Inside one provider, or above all of them.
| Tool | Memory Default | Cross-Model | Primary Use |
|---|---|---|---|
| ChatGPT | On by default | Export only | General assistant + memory |
| Claude | On for chat (Mar 2026) | Import + export | General assistant + transparent memory |
| Gemini | Account-linked, opt-in | Import + export | Google-integrated assistant |
| Grok | User-controlled | No cross-provider | Assistant with vector memory |
| Perplexity Spaces | Per workspace | Within Perplexity | Research workspaces |
| ReGild | Per persona, encrypted | Persona portable across providers (BYOK) | Persona that travels across models |
For the architectural detail on persona portability, see What Persona Portability Actually Means. For how ReGild handles your data while it does this, see our security page.