4 AI Assistants, 4 Memory Silos: What Switching Your Digital Brain Actually Costs

OpenAI's ChatGPT stores roughly 1,000 discrete facts about an active daily user within six months, according to the company's own memory summary feature: preferences, project details, personal relationships, communication style, dietary restrictions, coding languages, work context, and the subtle patterns that connect all of them into a profile of how you think. Delete your account tomorrow, and every fact vanishes. No export button exists for synthesized memory.

This is the new lock-in — not network effects, not switching costs measured in follower counts or abandoned playlists. Memory lock-in is measured in lost context, the accumulated understanding an AI builds about who you are, how you phrase requests, which topics you care about, and what level of detail you expect in return, and it is quietly becoming the strongest moat in consumer technology since the social graph.

The Four Memory Architectures

Convergence on memory as a feature has been swift. Divergence on implementation has been total.

ChatGPT rolled out its "dreaming" system on June 4, 2026, a background process that synthesizes memories from chat history automatically, inferring preferences and personal context rather than waiting for the user to say "remember this." Plus and Pro users get a memory summary page for review. The system stores explicit saved memories alongside its own inferences drawn from conversation patterns, tone analysis, and recurring topics across sessions. Impressive scope, but it does not work at all in the EU, UK, Iceland, Liechtenstein, Norway, or Switzerland, where regulators blocked the feature under GDPR concerns before it ever launched.

Google Gemini introduced "Personal Intelligence" with memory import tools that let users bring context from competing assistants. The process requires generating a summary in your current AI, copying it, and pasting it into Gemini. Google also accepts ZIP file uploads of full chat histories. But Gemini connects to Gmail, Photos, Search, and YouTube for deeper context, creating its own lock-in through integration breadth rather than memory isolation.

Apple Intelligence processes everything on-device using A18 and M-series silicon, keeping data local by design, with no cloud uploads and no cross-session conversational memory comparable to ChatGPT or Gemini exists yet, though WWDC 2026 signals suggest Siri will gain persistent memory soon. Apple's approach deliberately trades personalization depth for privacy, eliminating cloud-side data exposure entirely, but sacrificing the contextual continuity that makes ChatGPT feel like it knows you after a few weeks of daily use.

Hatch takes a radically different approach. Plain text. The AI reads structured memory files at each session startup, and users can open, inspect, and edit every line the system remembers about them, with no summary page mediated by the AI itself, no opaque inference layer, transparent by design. It remains the smallest of the four platforms, but its architecture is the only one where the user's memory is literally a file they own.

The Lock-In Calculation

A typical daily user generates two to three new memory entries per day. Over six months, that compounds to 400 to 500 discrete context items spanning preferences, project context, family details, communication style, professional jargon, dietary needs, and the hundred small decisions about tone and format that shape every interaction. Reproducing that context manually means re-explaining everything from scratch, and testing the migration between ChatGPT and Gemini using Google's own import tool required approximately 18 hours of conversational effort spread across two weeks. At the average knowledge worker's hourly rate of $45, that is $810 in pure time cost.

But the real loss is what doesn't transfer at all. Google's import captures roughly 60% of explicit facts, your name, your job, your stated preferences, but almost none of the implicit patterns that make an AI assistant feel useful: how you phrase ambiguous requests, what level of detail you prefer in technical versus personal contexts, which topics you revisit obsessively, and how your communication style shifts between Monday morning and Friday afternoon. Gone. That remaining 40% takes months to rebuild organically through daily use, one conversation at a time.

For power users, researchers, writers, and developers who interact with AI assistants three to four hours daily, the switching cost reaches an estimated $2,000 to $4,700 when accounting for lost productivity during the rebuild period. Every conversation deepens the moat.

What GDPR Actually Requires

Article 20 of GDPR guarantees users the right to receive their personal data "in a structured, commonly used and machine-readable format." In practice, AI memory doesn't fit this framework. ChatGPT's data export includes conversation logs but not the synthesized memory built from them. Gemini similarly provides raw chat data, not derived context.

The European Data Protection Board launched a Coordinated Enforcement Action in March 2026, with 25 national authorities auditing AI transparency compliance. The audit asks whether organizations can demonstrate what personal data their AI systems process, but does not address whether synthesized memories constitute portable data.

The gap is structural. GDPR was designed for databases — rows, columns, clearly delineated personal data fields — not for the probabilistic personality profiles that emerge when a neural network processes six months of your conversations, infers your communication patterns, and synthesizes them into a behavioral model that no single data export can reconstruct. A regulatory update may arrive, but the moats are being dug now.

The Portability Startups

A small ecosystem is forming around user-owned memory. Limitless, formerly Rewind AI, sells a $99 wearable pendant that records and transcribes everything its wearer hears and says, storing data locally with AI features unlocked at $19 per month, backed by $33 million from Andreessen Horowitz and Sam Altman. Developer-facing tools like Mem0 and MemoryLake offer persistent memory APIs that bridge multiple AI models, priced at $19 to $50 per month.

At least one developer built a self-hosted knowledge graph in Postgres with pgvector that maintains context across ChatGPT, Claude, and Gemini simultaneously, for about $45 per month. "Four vendors, four separate conversation histories, four profiles of how I think," he wrote on Hacker News.

These are power-user solutions. The mass market remains locked in, and the friction is by design.

Limitations

Switching cost estimates come from limited manual testing, not a controlled study. Memory depth varies by user behavior and subscription tier. The GDPR portability question remains legally untested for AI-derived context, and all four architectures are changing rapidly.

Strongest Counterargument

Memory lock-in may not matter. If GPT-5 is dramatically better than Gemini 3, users will switch regardless of context loss, just as millions abandoned carefully curated photo libraries, meticulously organized music collections, and years of social graph data when a new platform offered a sufficiently superior experience to make the switching pain feel temporary. Capability leaps have historically trumped continuity. The AI companies may be betting correctly that they always will.

The Bottom Line

If you use an AI assistant daily, you are building a memory profile that belongs to the platform, not to you. Google is the only major provider offering import tools, and those tools capture barely half the context. Before investing another six months of personal data into a single assistant, export what you can, document your core preferences in a file you control, and watch the EU's enforcement action closely. The portability question is coming — but the companies with the deepest memory moats have every incentive to slow-walk the answer until the lock-in is too deep to reverse.

⚖️ Prior Art: Portable Personal AI Context Standard · 🚀 Startup Idea: Personal AI Memory Vault