Auto-Generated Memories: Sapience Learns About You While You Work
Once a day, Sapience reads the messages you have sent and saves the new durable facts it finds as Agent Memories - short, dated, traceable, tagged, and never a password or API key. Opt-in, emailed to you each morning, and always in your control.
Overview
Every agent in Sapience works better when it already knows the basics about you - your role, your organization, the clients you talk about, the way you like things done. Those facts live in your Agent Memories, and until now the only way they got there was you (or an agent you asked directly) writing them down.
Auto-Generated Memories does that curation for you. Once a day, Sapience re-reads the messages you have sent since its last pass, picks out what is worth remembering, and saves each one as a memory - exactly as if you had asked an agent to remember it. The next agent you talk to starts the conversation already knowing them. If anything new was saved, you get an email the same morning listing it.
The feature is opt-in, per user. Nothing changes for you until it is switched on for your account.
What counts as a memory?
Two questions decide it: is this about your world - you, your people, your organization, your standing arrangements - and will it still be worth knowing later?
Saved as a memory | Not saved |
"Ken is the managing director of Main Beach Advisors." | "Summarise this document for me." (a request, not a fact) |
"Prefers board papers in Australian English." | "Was frustrated with the printer this morning." (transient) |
"Works with the Acme Industries account." | Anything true only inside one document, meeting, or conversation |
The second question is the one that trips people up. A durable fact about someone in your world still counts even if it first came up while you were researching something else - what matters is whether it is about your world, not how it surfaced. What is excluded is material you are merely processing: a pasted board pack, an attached report, a forwarded email. Those are the agent's raw material for today's task, not facts about you.
Secrets are never saved
Passwords, API keys, access tokens, private keys, connection strings, one-time codes, card and account numbers are never recorded as a memory - not in the text, not in the tags, and not even when you typed them yourself and they would technically be "durable and useful".
The reason is worth understanding, because it is the opposite of how storage usually works. A memory is not filed away and left alone - it is handed to your agent at the start of every future conversation. So a saved credential would not be a secret kept; it would be a secret repeated, indefinitely, in every session. Sapience will happily remember that an account or system exists and what you use it for. It will never remember the thing that opens it.
The four kinds of memory
Not everything worth remembering is permanent, so each memory is filed under one of four kinds:
Kind | What it holds | Example |
Fact | Something that stays true - you, your people, your organization, your tools | "ACME is headquartered in Queensland, Australia." |
Preference | How you like things done - tone, format, tools, ways of working | "Prefers short answers with the recommendation first." |
Instruction | A standing directive to your agents | "Always cite the source when quoting legislation." |
Activity | What you were doing or looking into that day - true then, not forever | "Reviewing the Acme supply agreement for renewal terms." |
Activity is the newest of the four, and it is kept deliberately separate. It gives an agent useful recent context - what you have been working on - without that context being mistaken for a permanent truth about you. Your agents read the durable kinds first, so a busy week of research never crowds out what they know about you.
Every memory is short, dated, and traceable
Three things are true of every memory Sapience writes for you:
- Short. One sentence about one thing. A message that carries two facts worth keeping becomes two memories rather than one long one.
- Dated. Each memory carries the date of the conversation it came from, not the date it was filed - so a fact picked up from a chat last March is plainly older than one saved this morning.
- Traceable. Each memory records the conversation behind it, so you can open that conversation and read what was actually said.
Memories also arrive tagged with a few topic words - People, Clients, Travel, Finance - so the Agent Memories screen filters the same way whether a memory came from the overnight pass or from an agent you asked to remember something.
Your morning email
When the overnight pass saves new memories for you, Sapience emails you a list of them, each with its kind and the date of the conversation it came from. If nothing new was worth remembering, no email is sent - quiet nights stay quiet.
The email is a summary, not a place to act. Everything in it is already in your Agent Memories screen, which is where you edit or delete anything you would rather it had not kept.
Where the memories appear
Generated memories land in the same Agent Memories screen as the ones you create yourself, and they behave the same way - every agent you talk to can draw on them.
Each one is marked as agent-created, so you can always tell which memories Sapience wrote for you. You stay in control: review them, edit the wording, or delete any you don't want, exactly as with a memory you added by hand.
What Happens Behind the Scenes
- Once a day, Sapience checks which users have Auto-Generated Memories switched on.
- For each enrolled user, it gathers only the new messages - the ones you have sent since the last daily pass. Nothing is re-read twice.
- A specialist extraction agent reads those messages alongside your existing memories, so it doesn't save something you already have. If something has changed - a new title, a new client - it saves the updated fact.
- For each thing worth keeping, the agent points at the specific message it came from. Sapience takes the date and the conversation from that message itself rather than asking the agent to supply them, so what you see on a memory is the real record, not a guess.
- Each new memory lands in your Agent Memories screen, marked as agent-created, filed under one of the four kinds, and tagged.
- If anything was saved, you get the email. If nothing durable turned up that day, nothing is saved and no email goes out. Quiet days produce no clutter.
Each user's messages are processed separately and privately - the extraction only ever sees your own sent messages and your own memories.
Turning it on
Auto-Generated Memories is being rolled out as an opt-in setting and is off by default. If you would like it enabled for your account, contact your Sapience administrator or support - there is nothing you need to configure beyond saying yes.
Summary
- Once a day, Sapience reads the messages you have sent since its last pass and saves what is worth remembering about your world as memories.
- Each memory is one short sentence, filed as a fact, preference, instruction or activity, tagged, dated with the conversation it came from, and traceable back to it.
- If anything was saved, you get an email that morning listing it.
- Memories appear in the normal Agent Memories screen, marked as agent-created, and you can edit or delete them freely.
- It is opt-in per user and off by default; nothing changes until it is enabled for your account.
- Never saved: passwords, API keys, access tokens, private keys, connection strings, one-time codes, card and account numbers.