Nobody signs up for lock-in. It accumulates. You pick an AI tool because it’s good at the task in front of you this week, you invest a few months of usage into it — custom instructions, project context, the accumulated back-and-forth that made it start giving useful answers instead of generic ones — and by the time a better option shows up, switching costs more than staying, even if staying is worse.
That’s the actual mechanism. Lock-in isn’t a contract clause. It’s the gap between what you’ve invested in a tool and what you’d have to redo to leave it.
Where the cost actually lives
Re-teaching context is the biggest one. Every AI tool with any kind of memory or personalization has spent weeks or months learning your codebase’s quirks, your team’s terminology, your product’s history. None of that transfers. A new tool starts at zero, and “starts at zero” for something you’d normally onboard a new hire with over their first month is a real cost, paid every time you evaluate an alternative — which is exactly why people stop evaluating alternatives.
Export is a checkbox feature, not a real capability. Plenty of platforms let you download a JSON dump of your data. Far fewer let you take your memory and actually use it somewhere else — in a format another tool can ingest, structured the way another tool expects. Export that produces a file nobody can import isn’t portability, it’s a compliance answer to “can I get my data out.”
The switching cost grows precisely when you’d most want to switch. The tools worth evaluating are the ones you’d use heavily — which means the accumulated context you’d be walking away from is largest exactly when the case for leaving is strongest. That’s not a coincidence; it’s the retention mechanism working as designed.
What to actually check before you commit
- Is memory addressable through an open protocol, or only through that vendor’s proprietary API? MCP-based memory can be called by any MCP client, which means the vendor isn’t the only door in or out.
- Can you inspect the underlying storage, not just export a snapshot? If you can query it directly, you’re not dependent on the vendor building an export feature you happen to need.
- Does switching clients require re-onboarding, or does existing memory carry over? This is the test that actually matters — not “can I get a data dump,” but “if I start using a different AI tool tomorrow, does it already know what the first one knew.”
The alternative isn’t “avoid AI tools”
None of this is an argument against investing in AI tooling — it’s an argument for investing in AI tooling whose value compounds independently of any single vendor staying good, staying priced the way it’s priced today, or staying in business at all. Memory that lives in an open, portable layer means the tool you switch to six months from now inherits the value you built up with the tool you’re using today, instead of resetting it. That’s the difference between lock-in and an actual asset.
That design target is spelled out in Why Your AI Memories Should Be Portable. Native product memory versus a shared layer is compared in ChatGPT Memory vs. Claude Memory vs. a Shared MCP Memory Layer.
Start your 14-day Nexus-Catalyst trial to keep project context independent of any one vendor.