What Ezra knows about you
How memory works in plain English. How to see what he knows. How to correct it. How to delete everything.
The short version
Ezra remembers things that help him be useful: your name, your work, your basic preferences, the people you mention often, patterns you confirm. He doesn't remember the contents of your emails or calendar events. He doesn't browse your stuff in the background. You can see everything he knows by texting him "what do you know about me," and you can delete any or all of it any time.
Three commands to know
"What do you know about me?"
Ezra responds with a plain-English summary of his profile of you. Things like:
The list is editable. If anything is wrong or out of date, just say so.
"Forget [X]"
Targeted deletion. Examples:
- "Forget that I'm vegetarian" — removes the dietary preference
- "Forget that conversation about the Madison house" — removes that episodic memory
- "Forget the recruiter pattern" — removes a learned behavior
Ezra confirms what he's removing before he removes it.
"Delete me"
The nuclear option. Wipes your entire account and everything associated with it. Within 24 hours, every database row is gone. Within 30 days, it's removed from backups too. After that, no recovery is possible — by design.
Ezra confirms before doing this. (Some people text "delete me" by accident — he checks.)
What's in the memory, in detail
Three layers, each with different rules.
Layer 1: Profile (your basic facts)
Long-term things that don't change much. Your name, your role, your timezone, the names of people you mention often, your basic preferences. Updated when you correct them. Deleted when you say so. Doesn't decay automatically.
Layer 2: Episodic memory (specific things that happened)
Recent conversations and decisions. "Approved draft to Sarah on March 5." "Mentioned planning daughter's birthday party on June 12." "Rejected the auto-archive suggestion."
Default retention: 90 days, then automatic deletion. You can change it:
- "Keep memory for a year" — extends to 12 months
- "Keep memory for 24 hours" — minimal retention, mostly fresh-start
Layer 3: Learned patterns (rules)
Patterns Ezra notices in your behavior, but only after he's surfaced them and you've confirmed. "User always declines recruiter emails." "User reschedules Friday meetings." Each one comes with provenance — Ezra can show you the events that led him to learn it, and you can revoke any pattern at any time.
What's not in the memory
- The contents of emails Ezra reads on your behalf (he processes and discards)
- The contents of calendar events (same)
- The contents of Notion pages, Slack messages, etc. (same)
- Voice recordings (we don't have any)
- Biometric data (we don't collect any)
- Browsing history outside the service
- Anything from connected services beyond what you authorized
The principle: store the smallest amount of state that makes the next interaction useful.
Memory is not training
People sometimes worry that "Ezra learns from you" means their data gets baked into the AI model that other users share. It doesn't.
The AI model — Anthropic's Claude — never trains on your data. We use the API explicitly configured to not contribute to training, and we never opt in to training. The model your Ezra uses today is the same model every other Ezra user uses. It doesn't have your patterns, your preferences, or your data baked into it.
What's personalized is the context Claude sees on each call, not the model itself. Your patterns + preferences are sent in along with your message, used to generate the response, and then discarded for that turn. Claude doesn't remember between calls. Ezra remembers — by storing your context separately and re-sending it each time.
This means: your data never leaks into someone else's Ezra. Your patterns never improve someone else's product. Your privacy is bounded by your account.
Memory you can audit
The full memory log is available any time. Text "show me my memory log" and Ezra exports a complete record of everything he's ever stored about you, when he stored it, and why.
You can also subscribe to memory updates: "let me know when you store something new about me" — Ezra will text you a brief notification any time something is added to his profile of you.
The bottom line
Memory is what makes Ezra feel like he knows you. It's also a place where things can go wrong if you can't see, edit, or delete it. So we built it to be transparent: the entire memory is something you can ask about, modify, or destroy with a single text.
If anything ever feels off — if Ezra knows something you didn't mean for him to know, or got something wrong, or surfaced something you'd rather he forget — just tell him. The default response is "got it, fixed."