Trevin Chow spent months getting his own backup of Google Photos. Takeout technically works, but nearly a terabyte arrived as hundreds of zip files, one at a time, through protection that treats automation as an attack. Getting data in is easy; getting it out is adversarial by design.
His argument is that personal AI agents are heading the same way, and that what they hold is as valuable as thirty years of photos.
The export is a snapshot, not a move
Every agent he uses keeps its own memory: what he's working on, how he likes things done, where he left off. The platforms vary in whether they let you export. Claude's import is essentially a prompt you paste in.
That's the crux: "That's a one-time snapshot, and the day after you copy it over the two agents start drifting apart again, because each keeps learning things the other never hears about. You haven't moved your memory. You've forked it."
A fork is not portability. Portability means the state moves; a fork means two divergent copies, and the longer you use both, the less either resembles the whole. For documents, that's annoying. For an assistant that knows your preferences and half-finished projects, it's a switching cost that grows every day you stay.
The second-brain itch, and why it came back
Chow notes the arc: people were briefly obsessed with "second brain" tools — gBrain, Supermemory, Hindsight, Honcho — and then the current crop of hosted agents arrived with memory built in, and the itch seemed to go away.
He self-hosts gBrain and describes it as a maintenance nightmare: broken ingestion, embedding warnings, failed "dream cycles," errors most weeks. His conclusion is blunt: "No one really wants to be an IT admin."
So the hosted agents won by default, not by solving the problem. They absorbed the feature and kept the data. The promise of a second brain — memory that belongs to you and travels with you — is still unfulfilled; it's just been relocated inside each vendor.
The lock-in is the point
None of these platforms is incentivized to make it easy to leave, or to help another agent pick up where they left off. That's not a conspiracy; it's the normal economics of switching costs. The more an agent knows about you, the more valuable it is, and the more expensive it is to replace.
Chow's own usage split — 60% one bot, 35% another, the rest dribbling — illustrates the trap. He has no loyalty to any of them, and he still can't switch cheaply, because switching means re-explaining everything.
"Do that enough and you feel friction to stop switching, which is the lock-in working exactly as they intend."
What would actually fix it
His bet: memory that sits outside any single agent, somewhere they can all read from and write to, so the memory is yours and the agent is whoever you're talking to that day.
That's the right shape, and it's harder than it sounds for three reasons the post doesn't dwell on.
Consistency. A shared store that multiple agents write to needs conflict resolution, provenance, and a notion of what's current. This is a distributed-systems problem wearing a personal-assistant costume.
Trust. A memory layer readable by every agent is a shared attack surface. If an agent can write to it, a prompt injection can write to it, and then the poisoned memory follows you into every future conversation. The read side is a privacy question; the write side is a security one.
Standards. Data portability worked because the format was simple: files and folders. Memory is not a file format; it's a set of claims about a person, with different confidence levels and different sensitivity. A standard would have to model that, not just serialize it.
None of which makes the goal wrong. It makes it a protocol problem, and protocols need either a dominant implementer or a regulator.
The willing-hostage framing
Chow ends with "willing hostages, but still hostages," and that's the honest version. The convenience is real. The memory genuinely helps. The lock-in is the price, and most of us pay it because the alternative is being an IT admin for a system that breaks weekly.
The uncomfortable parallel is the one he opens with: Google Photos. We accepted that getting our data out would be painful, and then we built a new layer of the same problem on top, with something more intimate than photos.
The question isn't whether agents should remember. It's who owns the memory, and whether the answer survives the next product launch.
Source: Are our personal AI agents holding us hostage? — Trevin Chow.