Memory & entities

Memory and entities are the same thing. Personal Agent keeps a single layer of knowledge about your world, and it surfaces on one screen — the Knowledge page (in the account menu). There is no separate "memory" page and "entities" page: the old /memory and /entities links both redirect to /knowledge, because they were never two things to begin with.

This page explains that one layer: what it is, how the assistant learns it, how to inspect and correct it, and how live integration data fits into the very same graph.

One layer, two kinds of knowledge

Everything the assistant knows is stored as a graph of entities linked by facts. That graph holds two kinds of knowledge that most products keep apart — and here they are unified:

1. Durable memory — things the assistant has learned: that you prefer terse answers, that "the K8s migration" is a project, who a person is. These are entities and facts the assistant builds up over time as you chat.
2. Live entities — things with a current state, synced from your integrations: a light that is on, a sensor reading, a calendar, a GitHub pull request, a project in OpenProject.

They are not two registries. An integration entity can declare a memory kind, and the sync engine then keeps a graph node for that entity inside the same memory graph. So a single "Living-room lamp" is both a live-state thing you can switch and a memory node that facts can attach to ("installed by Alice", "part of the evening scene"). On the Knowledge page it shows a Live badge with its current state, sitting in the same list as purely-learned entities like a person or a preference.

Why this matters

Because they share one graph, the assistant can reason across both at once — "turn off the lamp Alice installed" joins a learned fact (who installed it) with live state (which lamp, is it on). Ask it about a project and it can pull both what you've told it and the live issues synced from your tracker.

What the Knowledge page shows

The Knowledge page is one list of everything the assistant knows about your world — people, places, preferences, tasks, projects, devices, lights, sensors and more. Each row is an entity. Entities fed by an integration carry a Live badge and their current state; purely-learned entities don't.

You can:

• Search the graph and filter by kind (person, task, place, light, …) and by integration.
• Filter by nature — every kind is tagged physical (a lamp, a person, a room) or virtual (a project, a preference, a note topic).
• Switch between a list view and a graph view. The graph draws entities as nodes and their relationships as edges, and updates live as new facts are committed. It shows up to a few hundred nodes — filter or search to focus it.
• Open any entity for its detail page (covered below).

contact is the single kind used for people, so a person you email, message and meet is one entity — not one per channel. See Contacts for the person-centric view built on top of it.

Entities in detail

Open an entity to see and manage everything about it:

• Facts — every statement where this entity is the subject (preference = terse, location = Kitchen, state = on). A fact whose object is another entity is a relationship and shows under its own heading.
• Live state & attributes — for integration-linked entities, the current value and the raw attributes the integration reports.
• History — how the entity's facts and state have changed over time.
• Cause and effect — a trace of what led to a change: the run, message or sync that caused it, and what it resulted in.
• Time travel (see below).
• Forget — remove the entity from active memory.

Creating your own entities (helpers)

You don't only get entities from integrations — you can create your own, called helpers, right on the Knowledge page. A helper is a small virtual entity the assistant (and your dashboards and workflows) can read and control:

Helper	What it is
Toggle	An on/off switch
Number	A value with a min/max/step (rendered as a slider)
Button / counter	A press button or a +/- counter
Select	A dropdown of options you define
Text	A free-text value
Date / time	A date and/or time value

Helpers with the same device group are shown together under one device.

Areas, floors and devices

Entities can be organized like a real building: assign each to an area (a room), group areas onto floors, and roll physical entities up under devices. Areas and floors are managed from the Knowledge page and drive area/device cards and per-room dashboards.

Sync and search indexing

Integration entities refresh automatically; Sync now pulls the latest state on demand. Many entities are also indexed for semantic search (RAG) — marked with a RAG badge — so the assistant can find a project, document or issue by meaning, not just by name.

How the assistant learns

You rarely add memory by hand (though you can ask the assistant to remember a specific thing). The pipeline keeps writers deliberate and the chat agent read-only:

1. The chat agent only reads memory during a conversation.
2. After a run, a background curator reviews what happened and proposes memory updates — it never writes directly.
3. Clear, low-risk facts (things you stated or the assistant plainly observed) are committed automatically. Anything inferred, privileged or drawn from untrusted content is held as a suggestion to confirm.

Pending suggestions appear at the top of the Knowledge page, where you Keep or Reject each one. Conflicts and merge suggestions also surface conversationally in your main chat.

Time travel

Memory tracks two senses of time for every fact, so it can answer both "what was true on Monday?" and "what did you know on Monday?" — and explain late knowledge or corrections. The time-travel control on an entity lets you pick a past moment and see exactly what the assistant knew then. It's the quickest way to understand how a conclusion was reached or to audit a change.

Correcting and forgetting

Knowledge changes, so memory is built to be corrected without losing its past:

• Editing a fact supersedes the old one rather than erasing it, so history and time-travel stay intact.
• Forget removes an entity from active memory.
• When an integration stops reporting an entity, its node is archived, not deleted — so any facts you attached to it survive, and it can come back if the integration does.

Privacy and per-chat access

Memory is strictly private to you, isolated at the database level. On top of that, each chat decides how much memory it may read with the Memory control — Full, None, or Restricted to chosen areas and sources. A None chat is a clean room: nothing is read and nothing is learned from it. See Chat controls.

The simple remembered-facts list

For a plain, flat view of the facts the assistant has stored about you — the ones it picked up with its remember tool — open Settings → Memory. It lists every entry, marks whether it came from you or from the assistant, and lets you delete any of them. It's a lightweight companion to the full Knowledge page above.

Going deeper

The complete design of the world-state memory graph — its two time axes, provenance, and the read / propose / write pipeline — is documented in Universal memory.