Agents discover. You confirm. That is how personal memory becomes trustworthy.

The smartest model still cannot guess your life

Large language models are excellent at public knowledge: facts that are broadly true and can be checked by many people. Personal facts are different. No model can infer, from the internet, how you want to be addressed, why you rejected a plan last year, which tool you prefer for a recurring job, or which boundary an agent must never cross.

A stronger model may make a more plausible guess. That is precisely the danger: a polished wrong answer about you can sound completely reasonable, and only you can correct it. Better retrieval does not solve an empty foundation. If the information was never captured, there is nothing to retrieve.

You should not have to write a user manual about yourself

The usual answer is a profile, custom-instructions box, or hand-maintained memory file. These are useful escape hatches, but they are a poor main path. Familiar facts feel like background to you, so you do not know which ones are new to an agent. The highest-value memories — the reason behind a decision, a lesson from a failure, a quiet working preference — are also the hardest to summarize on demand.

And you change. A profile written once is a photograph; a person is a moving story. A memory product should not turn self-description into another inbox you have to maintain.

The useful facts are already in the work

You may never sit down to tell an AI, “this is how I make decisions.” Yet you reveal it naturally in project discussions, email, meeting transcripts, agent sessions, and the documents you edit. Those moments carry what a form loses: first-person language, time, audience, reasons, and surrounding context.

That suggests a better division of labor:

1. You bring selected work and conversations into your vault. The source remains available as evidence and stays under your control.
2. An agent discovers small, atomic candidates. It proposes one clear statement at a time instead of inventing a complete profile.
3. You judge each one. Confirm it, deny it, mark it important, or ignore it.
4. Only your decision creates trusted memory. Approved claims become durable, versioned assets that can produce simple USER.md and MEMORY.md views.

The machine does the scanning and drafting. You keep the one job that cannot be delegated: deciding whether a statement really represents you.

A recording is evidence, not automatically a fact

Everyday language is messy. “We will move to PostgreSQL next quarter” might be a decision, a proposal, a joke, a quotation, or an assumption in a thought experiment. A transcript may even attribute it to the wrong speaker. Flatten all of that into a sentence called a fact and the system throws away exactly what made the sentence trustworthy.

note.md therefore stores claims, not anonymous facts. A claim keeps who it is about, who asserted it, where it came from, when it applied, and what kind of approval it received. The source can help an agent propose a memory; it cannot approve the memory on your behalf.

This is why the review step is not temporary scaffolding to remove when models improve. It is the part that creates authority.

One Confirm button hides three different decisions

“Remember this,” “this external statement is true,” and “you may act on this” are not the same permission. note.md keeps them separate:

The words on the button change with the decision. Approval is bound to the exact content you reviewed; if the candidate changes, it must be reviewed again. Agents may propose, but they cannot manufacture a human approval.

Search and memory are different jobs

Many systems treat memory as a special search index, or treat the whole document library as memory. note.md deliberately separates the two.

Search gives you cognitive relief: you do not have to remember everything because you can find it. Memory gives agents cognitive alignment: they receive the few statements you have chosen to stand behind. Relevance cannot be used as a substitute for permission.

A real vault shows why the checkpoint matters

In a September 2, 2026 snapshot of one real note.md vault, agents had proposed 83 memory claims. The owner kept 27 and ignored 56. In other words, almost two thirds of plausible suggestions did not deserve to become lasting context.

This is not an argument that agents are unhelpful. It shows their best role. Nearly every candidate was discovered by an agent; the agent found valuable signals the person would never have written into a profile. The owner then removed the jokes, temporary states, repetitions, and overconfident interpretations. Discovery created coverage. Confirmation created trust.

What note.md promises you

• No guess silently becomes your profile. An agent proposal remains a proposal until you decide.
• You review one meaning at a time. There is no bulk approval for sensitive identity, boundary, or authorization claims.
• You can see where a memory came from. Source, author, time, approval, and revision history stay attached.
• A memory is used only for an allowed context. Space, purpose, provider, and sharing rules decide what an agent may receive.
• Your corrections compound. Ignored suggestions are suppressed, and guidance can record the error future agents must avoid.
• The asset remains yours. Authority lives in local, Git-tracked files; readable Markdown views can be rebuilt, and a different agent can use the same approved context.

Why the work gets lighter over time

Personal memory is not an endless stream of profile fields. Identity, stable preferences, and boundaries are relatively small sets. They are established early and then mostly revised. Decisions continue to arrive, but they can be reviewed as individual moments instead of forcing you to rewrite a biography.

The loop also improves itself. Approved claims help the next discovery pass avoid duplicates. Ignored candidates stop returning. Each correction teaches future agents what not to assume. The goal is not to ask you more questions; it is to reserve your attention for the few questions only you can answer.

Choose the memory contract you want

Different jobs deserve different memory contracts:

Choose note.md's approach if you would rather give an agent fewer memories you can trust than a larger profile built from silent guesses. It is designed for people who want useful personalization without handing an AI the right to define them.

The next improvements follow the same thesis: make discovery across user-selected sources more precise, pace reviews so they stay thoughtful, use time and actual recall to suggest when a memory deserves review, learn from successful and failed work, and publish repeatable quality measurements. These make the loop quieter and more helpful; they do not remove the human checkpoint.

An external database stores statements about you. A second memory earns your trust because you remember approving what it knows.