Notes on storage for personal memory

Early thinking

This is a note, not a result. We are writing down how we think about storage for personal memory, before we are sure.

Start with the wrong default. Today most "memory" lives inside the agent. Each assistant keeps its own profile of you, and none of them agree. That profile is the model's lossy summary of you, not a record you own.

Move the memory out. Treat a person's context as data with a schema, a home, and an owner — the person. The agent reads from it and writes back, but it is not the source of truth.

What the layer has to do

A few properties seem non-negotiable.

It has to be portable. Context trapped in one app is not a substrate; it is a silo.

It has to be legible. A person should be able to see what is stored and delete it.

It has to be cheap to read. If assembling context costs a second and a rupee per turn, agents will skip it.

The contribution is not better agent memory. It is moving memory out of the agent.

Open questions

• What is the right schema for "everything an agent should know about me" — and is one schema even possible?
• Where does the data live: on device, in a personal cloud, or somewhere shared?
• How do you grant and revoke an agent's access without a permissions dialog on every turn?

We do not have answers. We have a strong prior that the memory is in the wrong place.

@misc{datafrontier_data_infrastructure_note,
  title  = {Notes on storage for personal memory},
  author = {DataFrontier Team},
  year   = {2025},
  url    = {https://datafrontier.co/research/data-infrastructure-note}
}