When an assistant answers from your own notes, the stakes are different from ordinary chatbot errors. A wrong answer here rewrites your past: a commitment you never made, a decision you actually reversed. Trust in a personal AI is not a feeling—it is a checklist, and it has four items.
Visible sources
Every claim should point to the notes that support it—openable, with dates and excerpts. Sources do more than prove the answer; they scope it. Seeing that a conclusion rests on two notes from January tells you what the system does not know as clearly as what it does.
Honest chronology
Personal history is full of reversals: plans changed, quotes revised, opinions updated. A trustworthy answer respects sequence—quoting the March decision means also surfacing the June note that overturned it. The most recent relevant evidence defines the current state; everything older should be presented as history.
Assumptions labeled as assumptions
Sometimes the notes do not fully answer the question, and a helpful system bridges the gap with inference. That is fine—if the seam is visible.
- ‘Your notes say X’ must mean the notes actually say X.
- ‘Probably Y, based on…’ marks inference doing the work.
- ‘Your notes do not cover this’ is a complete and honest answer.
- Conflicting evidence is shown as a conflict, not silently resolved.
Correctability that sticks
You will catch mistakes—a misattributed quote, a stale conclusion. Trust survives only if the correction has somewhere to go: fix the record, and future answers reflect the fix. An assistant that repeats an error you already corrected has demonstrated that verification is pointless—and verification is the entire basis of the relationship.
How to test the four properties in ten minutes
Each property has a question that exposes it, and any tool worth a year of your material should survive all four.
- Sources: ask something you know the answer to, then check whether the cited notes are the ones you were thinking of.
- Chronology: ask about a decision you later reversed. A trustworthy answer surfaces the reversal; a weak one quotes whichever note matched best.
- Assumptions: ask about something you never wrote down. The correct answer is ‘nothing saved about that’, not a plausible construction.
- Correctability: correct one wrong detail, then ask the same question again a day later and see whether the fix held.
Why fluency makes this harder than it sounds
The specific danger with answers about your own life is that you have no external check. When a model is wrong about a public fact, the world contradicts it. When it is wrong about what you promised in March, the only thing available to contradict it is a memory you already know is unreliable—which is why you asked.
Worse, a confident wrong answer often feels more correct than a hedged right one. Reading ‘you agreed to send it by the 14th’ tends to produce a vague sense of recognition whether or not it happened, because recognition is easy to manufacture and hard to distinguish from recall. The sources are not a courtesy feature; they are the only defense against your own agreeableness.
None of the four properties makes an answer true. They make it checkable, which is a weaker and much more useful guarantee. The remaining work—opening the note, reading the passage, deciding whether the interpretation is fair—stays yours, and any product that suggests otherwise is selling confidence rather than memory.
Before relying on any answer about your own history, run the checklist: can I see the sources, is the newest evidence on top, are guesses labeled, and will my correction stick? Four yeses make a memory. Anything less is a fluent guess.