Memory Model
Ajeris is designed to remember the things about you that matter and forget the rest. This page describes how that works in practice so you know what to expect, what's worth saying explicitly, and what the agent picks up on its own.
Two ways memory gets saved
1. Explicitly, when you ask. Anything you say in the form "remember that X" or "from now on, do Y" is captured directly. This is the strongest signal and the agent treats it as a high-confidence instruction.
2. Passively, in the background. Every so often during a conversation, the agent quietly reviews what you've said and saves things that seem worth remembering for next time. You don't see this happen. It runs after the agent's reply has already been delivered, so it never delays your messages.
The passive review is conservative. Most of the time it concludes "nothing worth saving" and moves on. It only saves things when there's a clear signal.
What gets saved
The agent tries to capture three categories of information:
- Facts about you and the people in your life. Names, birthdays, addresses, work, family relationships, allergies, anything that would be useful as context in future conversations.
- Preferences and behavioral expectations. Communication style ("don't start replies with 'Got it'"), tone, level of detail, what you want the agent to ask before doing vs. just do.
- Routines. Recurring patterns it sees you do ("Uber to the airport at 6 a.m. on weekdays", "Friday afternoon walk").
If you tell the agent something once, it's likely to stick. If you mention something in passing that doesn't recur, it may or may not make it into long-term memory.
What does not get saved
- The full text of your conversations. Memory is a curated set of facts and preferences, not a transcript. Old messages are not recalled verbatim across sessions.
- Sensitive credentials or one-time codes. The agent specifically avoids capturing tokens, passwords, or codes you paste in.
- Things you ask it to forget. "Forget that I ever mentioned X" removes the corresponding memory.
Reviewing and editing memory
You can ask the agent at any time:
- "What do you remember about me?"
- "What have you saved recently?"
- "Forget that I said X."
- "Update what you have for my home address."
The agent treats these as direct edits to its memory store and applies them immediately.
How memory shapes future replies
When you start a new conversation, the agent loads the relevant slice of your memory before it reads your message. That's what lets it greet you correctly, know which devices and services you've connected, remember your preferences for length and tone, and skip questions it's already asked.
Memory is also what makes corrections stick. If you tell the agent once that you don't like a particular phrasing, every subsequent session inherits that correction without you having to re-state it.
Privacy
Memories are stored in a database encrypted at rest, scoped strictly to your account by row-level security, and never shared between users. See the security policy for the full data-handling details.