Introducing Memory & Personalization: AI That Knows You
You open a new chat. You ask it to draft an email. It writes something generic, polished, and completely wrong for your voice. So you correct it. Again. For the fourth time this week.
Or you start a fresh thread with a different agent and have to re-explain your role, your project, your preferences. Last conversation's context evaporated the moment you closed the tab.
Memory & Personalization ends both of those problems. Your AI knows who you are, what you're working on, and how you like things done. That knowledge carries into every thread, with every system agent, from the first message.
And unlike most AI memory, this one is a glass box. You can see the exact context agents receive. You can keep, edit, hide, or delete any fact. You can turn the whole layer off.
Start here: Settings → Personalization
Open Settings → Personalization. That page is the control surface. Everything below lives there.
1. The master toggle
At the top of Personalization there is an On / Off switch.
- On (the default): agents receive your profile and active memories. New conversations can teach the system.
- Off: agents get no personalization envelope. Nothing new is extracted. Stored facts stay in your account until you delete them; they just stop being used.
Flip the switch, then tap Save profile. The toggle does not persist until you save.
Use this when you want a blank-slate thread, when you are handling something you do not want remembered, or when you simply want the layer quiet.
2. Tell it who you are
Fill in the identity fields in your own words: preferred name, role, what you're working on, how answers should feel, goals, hard preferences.
Or tap Set up with questions. Six short prompts, then a review card: you see what was derived, keep what is true, discard the rest. Nothing from that flow reaches agents until you approve it.
This is the cold-start fix. A brand-new thread already knows your name, your work, and your constraints.
3. Approve what we already learned
While we were building this, ARMES extracted durable facts from conversations and held them for you. Those sit in Needs review.
Each card shows:
- The fact, in plain language
- The evidence quote (the words it came from)
- A trust badge: You said this, From a conversation, From your history, Inferred, and similar
- Keep or Discard
Until you Keep a proposed fact, agents do not use it. Disputed items (an inferred fact colliding with something you typed) and orphaned items (the source chat was deleted) also wait here.
Work through the queue when you have a few minutes. Keep what is true. Discard guesses. That catch-up is a one-time pass, not a new chore for every chat.
4. Going forward: automatic
From here on, durable facts picked up in new conversations can activate on their own. You do not have to approve every "I use Cursor" or "keep answers short."
You still own the layer:
- Every activated fact appears on the In use tab
- You can edit the wording, pin it (always included), hide it from everyday context ("only when I bring it up"), or delete it
- Sensitive categories are never auto-extracted. Medical, political, financial, and similar content is not ambient, even if you later ask an agent to remember something in that category
- A lower-trust guess can never silently replace something you typed in Settings. Conflicts come back to Needs review for you to settle
Automatic means useful continuity. It does not mean a hidden profile. What agents see at the bottom of the page is the exact block injected into conversations. If it is not in that preview, they do not have it.
5. Backfill from history (opt-in)
If you have older chats you want folded in, scroll to Build from your history.
Tap Review my history. ARMES reads your longest and most recent conversations, writes a short searchable summary of each, and proposes individual facts. This is consent-gated. It is not on unless you start it.
Nothing from backfill becomes active until you Keep it in Needs review. You can leave the page while it runs. When it finishes, proposed facts are waiting in the queue. Check for new conversations later if you want another pass.
Backfill is the way to teach ARMES about work you already did, without re-explaining it in a new thread.
6. Conversation summaries
Below the ledger, conversation summaries (episodes) are rolling recaps of threads: what was decided, what is still open. Agents can search them. You can read and delete any summary. Delete a chat and its summary goes with it.
These are not a second secret memory. They are the retrieval unit that lets an agent find "we held pricing at $19" without rereading forty messages.
Talk to your agents
You do not have to open Settings for every change. In any 1:1 chat with a system agent:
- "Remember that I prefer X" — saved immediately. The agent confirms the wording it stored. No queue.
- "Forget that I use Y" — it finds the matching fact and removes it.
- "What do you know about me?" — a structured dossier: identity, goals, preferences, known facts, recent themes, and where to edit everything.
These tools share one ledger across agents. Ask your writing agent to remember a brand voice; your research agent will know it too.
Memory stays off in DMs and group chats with other people. Agents in those rooms do not build personalization from those messages. Custom webhook agents do not automatically receive the envelope.
What the system actually stores
Three pieces, all visible:
Identity. Name, role, current work, style, goals, hard preferences. You typed these, or you approved them from onboarding.
Facts. Atomic memories with a kind, a source, and a trust level. Pinned and ambient-eligible facts go into the envelope (capped so they do not crowd out the actual conversation). The rest stay searchable when you or an agent ask.
Episodes. Summaries of conversations, for search and for long threads that have outgrown the recent-turn window.
There is no dreaming layer, no advertising profile, and no "the summary page may not show everything." The envelope preview is the complete injection.
A worked example
You're a freelance brand strategist named Alex.
Day one. You open Personalization, leave the toggle on, fill in name, role, and "direct answers, no preamble." You Keep a handful of proposed facts from earlier chats. New threads already sound like you.
Later that week. You tap Review my history. After it finishes, you Keep "structures brand decks as Story, then Strategy, then System" and discard a stale client name.
A new conversation. You mention you always write in American English. That fact activates going forward. You do not have to approve it. You can still delete it from In use if it is wrong.
You change your mind. You tell an agent, "I've stopped using the three-S framework; lead with the audience's problem." The new preference supersedes the old one. History is preserved; agents use the current fact.
You need it gone. "Forget the three-S framework," or delete the row in Settings. After that, agents do not keep a shadow copy.
What makes this different
| Other AI memory | ARMES memory |
|---|---|
| Infers in the background; the summary may not show everything | Envelope preview: what you see is what agents get |
| You can view and delete, sometimes | You can view, edit, pin, hide from ambient context, and delete |
| One flat list | Structured facts with kinds, trust levels, and provenance |
| Silent overwrites when you change your mind | Never overwrites. New facts supersede old ones; history stays |
| A profile the vendor writes about you | A fact ledger you own, in your workspace |
| Provider-side retention and ads profiles in some products | Extraction runs on zero-data-retention inference. No ads profile. No training |
Catch-up and backfill wait for Keep. Going-forward conversation facts can activate automatically. Either way, the ledger is the source of truth.
Trust: not all facts are equal
ARMES ranks how a memory was created:
- Settings (highest): you typed it into your profile
- Onboarding: derived from answers you reviewed
- Remember: you asked an agent to save it
- Conversation: picked up from a chat going forward
- Backfill: extracted from history, still Keep/Discard
A lower-trust fact cannot silently replace a higher-trust one. If Settings says you prefer concise answers and a conversation suggests you want thorough explanations, the collision lands in Needs review. You settle it.
Privacy, in one paragraph
Personalization is on by default so agents can actually help. It is workspace data, not an ads profile and not a model the labs keep. Chat, extraction, summarization, and embeddings use zero-data-retention inference: providers process and forget; there is no human review of that path. You can turn the master toggle off, delete any fact or episode, or delete the source chat (proposed memories sourced only from it are removed; active ones may be flagged as orphaned for you to keep or drop). Memory is off when you are talking with other people. Detail: Privacy and Security.
Works with Projects
Global memories (your name, role, general preferences) apply everywhere. Project-scoped facts stay in that project. Switch workstreams and the wrong client's tone does not leak. Same isolation as notes and files.
Each project can also carry its own custom instructions, so the same agent adapts based on which workstream you're in.
On every plan, including Free
Memory & Personalization works on Free, Pro, and Ultra. A personal AI that actually knows you should not be locked behind a paywall.
The fastest path:
- Open Settings → Personalization
- Confirm the toggle is on and save if you changed anything
- Fill identity, or run the questions
- Clear Needs review
- Optionally Review my history
- Open a brand-new thread with any system agent
Or start free, no credit card required. Same memory, same control, on the house.
Your AI should know you. You should know what it knows.
Written by
ARMES Team
From the team building ARMES — private AI that puts every frontier model in one place.