Flagship report · 25 min read · July 2026 · updated August 26, 2026
A mid-year audit of what every major AI actually remembers about you — what changed in the last six months, what stayed exactly the same, and where the fight over ownership of your AI memory stands right now.
The one-paragraph summary
Every major AI provider shipped memory improvements in H1 2026. None of them shipped memory portability. Six months later, the situation is unchanged in the ways that most affect users: memory is deeper inside each provider's walled garden, and no easier to move between them. The interesting motion is happening in the middle layer — MCP adoption, third-party memory vaults, and a small but real class of AI-native power users who now treat "which model" as a monthly decision rather than an annual commitment.
Update · August 26, 2026
Six weeks after this report went out, the thing it described as slow motion stopped being slow. Memory is no longer a feature inside AI products — for most of the major providers it is now the feature they lead with. Everything below still holds. What changed is the pace, and the size of the gap between "our AI remembers you" and "you own that memory."
Anthropic, August 25. Claude now carries memory across chats and into Cowork sessions, with a readable, editable list of what it has retained, organised into topics you can delete individually. Sensitive subjects are off by default. It is the best consumer memory control anyone has shipped, and it deserves to be said plainly — we went through it in detail here. It is also, still, Anthropic's copy of you, on Anthropic's servers, in Anthropic's format.
OpenAI, free tier. ChatGPT's free plan now remembers substantially more, including a lighter version of the reference-past-chats behaviour that used to be paid-only. Genuinely good for the people it reaches — and more of you stored somewhere you don't control, which is a different thing from more that you own.
The memory layer got crowded. A visible cohort of startups now sells "memory for your AI" as a standalone product, and the platform players have started treating persistent context as a competitive axis rather than a convenience. They differ enormously in approach and share one structural trait: the memory lives inside somebody's product.
Portability. Not one of the August releases made it easier to move what an AI knows about you to a different AI. Memory got deeper, and the walls got taller at exactly the same rate.
Provenance. Every new memory surface shows you conclusions with the evidence stripped off. You can read "prefers direct feedback" and delete it, but you cannot find out whether it came from one irritable Tuesday or from a hundred consistent samples — and those deserve opposite treatment. We wrote up why an edit button doesn't close this.
Permanence. Model retirements kept happening on the providers' schedule, and still ship without a "here's what this model knew about you" export. Your stored list survives a sunset. The assistant that knew how to use it does not.
It would be easy to read a month of provider memory launches as the category closing. We read it the other way. Every provider that ships memory teaches millions of people to expect an AI that knows them — and the moment that expectation is normal, the next question arrives on its own the first time someone changes plans, switches tools, or watches a model get retired: whose copy is this?
Nobody shipped an answer to that in August. That question is the whole category, and it is still open.
Six product changes actually mattered in the first half of the year. Everything else was noise.
Twelve axes, six providers, one honest table. Scores are 1 (weakest) to 5 (strongest) based on our controlled testing across Q2 2026. None of the six scored above a 3 on portability. That's the story.
| Axis | ChatGPT | Claude | Gemini | Copilot | Perplexity | Grok |
|---|---|---|---|---|---|---|
| Explicit memory layer | 5 | 3 | 4 | 3 | 3 | 3 |
| Implicit context depth | 2 | 2 | 5 | 5 | 3 | 4 |
| Effective context window | 4 | 5 | 5 | 3 | 3 | 3 |
| Retrieval precision | 4 | 4 | 3 | 3 | 4 | 3 |
| Memory edit-ability | 4 | 4 | 4 | 3 | 4 | 3 |
| Cross-device continuity | 4 | 4 | 5 | 4 | 4 | 3 |
| Export completeness | 3 | 3 | 3 | 2 | 2 | 3 |
| Import from elsewhere | 1 | 1 | 1 | 1 | 1 | 1 |
| Model-update stability | 2 | 3 | 3 | 3 | 3 | 3 |
| Privacy controls | 4 | 4 | 4 | 5 | 4 | 3 |
| MCP support | 2 | 4 | 2 | 2 | 2 | 2 |
| Portability | 1 | 2 | 1 | 1 | 1 | 1 |
Every axis except one has at least one provider scoring a 4 or 5. Portability has no one above a 2. That is the entire structural story of 2026. Providers are competing hard on how much they remember and how well they retrieve — and not at all on letting you take that memory somewhere else.
The economic logic is obvious: memory is switching-cost, and switching cost is retention. A user who has spent six months training ChatGPT on their voice, preferences, and projects is much more expensive to poach than a user who's just signed up. Every provider knows this. None of them has a business incentive to make export-import easy across competitors.
What's changed in 2026 is that the user demand for portability has become visible. The rise of tools like Konshus, the (slow) MCP adoption, and the growing set of "help I want to switch from ChatGPT to Claude" search queries all point at the same underlying frustration.
Every model update in H1 2026 caused visible personality drift for a subset of users. GPT-5 was the most-discussed. Claude 4.5 was the most tonally-shifted. Gemini's summer update was the smoothest, largely because Gemini leans on live Workspace data, which is stable across model swaps in a way saved memories aren't.
The cost of a model swap for a heavy user is measurable: on controlled benchmarks (see persona drift, with data), tone and stated preferences drifted by 15–35% on the same input prompt across model versions of the same provider. That's not a subjective "feels different" — it's the underlying measurable behavior of the model changing under stable inputs.
The Model Context Protocol was supposed to be the portability primitive. In H1 2026 it's more like a portability primitive for developers, not consumers. Cursor, Cline, and other developer tools use it heavily. Consumer AI has been slower — Anthropic's native support is the standout; OpenAI has been quiet.
Our operator dashboard shows MCP client connections growing roughly 2× quarter-over-quarter in H1 2026, but the absolute numbers are still small — the modal AI user has never heard of MCP. The next inflection will be when a consumer AI product ships MCP-based memory injection as a default. Nobody has yet.
Given the landscape, the durable strategy for anyone who cares about continuity is unchanged since H1 and will likely be unchanged through H2:
The full step-by-step version of this playbook is in the complete guide to owning your AI consciousness.
We'll re-run the scorecard in January and score ourselves against these predictions. If you want to be nudged when the next edition drops, meet Konshus — you'll get a note when it publishes.
Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.
Meet KonshusKonshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.
Meet Konshus