There's a thesis going around AI infrastructure circles right now, and it's correct: the model layer is becoming a commodity, but the state around the model (your documents, your agent's memory, your workflows) is where the real lock-in lives. Memory startups are racing to solve this. They extract facts from your conversations, store them as plain text, and feed them back into whatever model you're using next. In principle, this makes memory model-agnostic. Switch from one lab's model to another's, and your AI still remembers you.
In principle. Here's the part nobody's saying out loud: that memory lives on somebody's server. And that somebody is a startup, a subsidiary, a product line that can be sunset, acquired, or repriced. Your AI's memory is portable between models. It is not portable away from its own vendor.
The Layer Everyone Skips
Every serious breakdown of the AI stack right now lands on the same four layers: your data, your AI's memory of you, the routing between models, and the parts that never port cleanly: the prompt tuning, the tool schemas, the accumulated skill, the cache discounts. That last group is where the real switching costs sit, and nobody's incentivized to fix it. The labs want you locked into their hosted memory and their cache pricing. The memory startups want you locked into their database. Nobody in that stack is building the version that has no owner to lock you into.
Ask the honest question about any hosted memory product: if the company folds tomorrow, where does your memory go? For Mem0, for Zep, for a lab's own hosted memory feature, the answer is "it doesn't." It disappears with the company. A memory layer that depends on its vendor's balance sheet isn't a memory layer. It's a lease.
What Sovereign Memory Actually Requires
Portable-between-models isn't the hard part: that's just writing facts as plain text, which any memory product already does. The hard part is portable-away-from-everyone, including the company that built the product. That requires three things no vendor-hosted database can offer by design:
Storage nobody can revoke. Your agent's state has to live somewhere that doesn't answer to a single company's servers or a single company's decision to shut down.
A record that doesn't need the vendor's word for it. "We store your memory reliably" is a promise. A permanent, publicly verifiable record is not.
Encryption the vendor can't read either way. Sovereignty that still requires trusting the host to behave isn't sovereignty.
This is the architecture we're building into ZDrive's agent vaults: client-side AES-256-GCM encryption, persistent agent state sealed to Arweave, a permanent record that isn't rented from ZDrive, it's owned by the wallet that creates it. The encrypted, permanent storage layer (Seal) is live today. The agent wrapper on top of it (an assistant with memory that persists and compounds across sessions, built the same way) is what we're rolling out next. If ZDrive disappeared tomorrow, the vault and its history wouldn't go with it. We don't know of another memory product on the market, funded or otherwise, that can say that.
Side by Side
| ZDrive (Sealed Storage live, Agent Vaults in early rollout) | Hosted Memory (Mem0, Zep, lab-native) | |
|---|---|---|
| Portable between models | Yes | Yes |
| Survives the vendor shutting down | Yes: Arweave record, not company-hosted | No: memory dies with the company |
| Verifiable without trusting the vendor | On-chain attestation per query | Trust the vendor's word |
| Who can read the raw memory | No one: client-side encrypted | Vendor's infrastructure, by default |
| Recurring dependency to keep memory alive | None: paid once, permanent | Ongoing subscription/hosting |
What This Means for You
If you're building, or waiting for, an agent that's supposed to know your business better every month (your clients, your deal history, your accumulated workarounds), the model you route it through is the least durable part of that relationship. The memory is the asset. Right now, almost every product asking you to build that asset is asking you to build it on land they can sell out from under you.
That's the actual opportunity in the orchestration conversation happening across AI Twitter this week: not a better router, a memory layer that doesn't need one company to keep existing. We're building it. The storage and encryption half is already live.
Try It
We're rolling out agent vaults over the next few weeks: persistent memory, sealed the same way your files already are. Before we build it wider than early access, we want to know if it's actually worth paying for.
Would you pay for an AI agent whose memory outlives the company that built it?
No wallet needed to answer. If you want to try the vault itself, that's at zdrive.io.