AI Memory
Admin → Insights → AI Memory shows everything the AI assistant (Ask GEM) has recorded as persistent knowledge from conversations — terminology, site conventions, device quirks, entity aliases, and workflows.
Open AI Memory
Why It Matters
This is the audit-and-correction surface for what the assistant "knows" about your specific install. The assistant writes these notes itself; this page is where you review them, fix anything wrong, and delete anything stale. If Ask GEM keeps repeating something incorrect or out of date, the bad note is almost always visible here.
Two things make these notes important:
- Site notes shape every answer. The most recent site-scoped notes are loaded into the assistant's context at the start of every conversation (see How notes reach the assistant below), so they directly influence its answers and the actions it takes.
- You teach it once. When you tell the assistant something about how this installation works — "we call the main receiver the brain", "macros 200–299 are test macros, never run them in production", "device 42 has a 30-second warmup, wait before sending commands" — it records that here and carries it forward across all later sessions.
Where these notes come from
You do not create notes on this page (there is no Add button). The assistant records a note on its own whenever you teach it something about this install that isn't already captured in the data model. Each note keeps the question that produced it, so you can see what prompted it.
If the assistant is later taught the same thing again (same Subject and same Scope), it updates the existing note rather than creating a duplicate — so you generally see one row per fact, not a growing pile.
To correct a memory, edit it here. To make the assistant forget something, delete it.
The Grid
The grid is titled AI Memory Notes and lists one row per recorded note.
| Column | Meaning |
|---|---|
| ID | Internal note number. |
| Scope | site — applies to everyone on this install; user — personal to the person who taught it. |
| Category | Free-form tag the assistant assigns, e.g. terminology, site_convention, quirk, entity_alias, workflow (any text is allowed). |
| Subject | Short headline of the memory. |
| Note | The memory content (truncated past ~80 characters in the grid — use the View icon for the full text). |
| Tags | Keyword tags the assistant uses when it searches its own memory. |
| Entity Type | Optional GEM entity this note relates to (device, zone, macro, …). |
| Entity ID | The id of that related entity, if any. |
| Learned | When the note was recorded. |
| Last Recalled | The last time the assistant searched its memory and this note matched — see the note below; this is not the last time the note was used. |
Last Recalled updates only when the assistant runs an explicit memory search and this note comes back as a match. It does not update when a site note is automatically loaded into the conversation context. Because site notes are auto-loaded on every session, a heavily-used site note can still show a blank or old Last Recalled. Treat an empty/old value as "the assistant hasn't had to search for it," not "nobody uses it" — read the Note text before deleting.
Actions
- View (eye icon) — opens the full note: subject, scope · category, when it was learned, the complete note text, and the original user question that produced it.
- Inline edit — click a cell to edit it. Scope (a
site/userdropdown), Category, Subject, and Note are editable directly in the grid so you can correct a memory. Tags, Entity Type, Entity ID, Learned, and Last Recalled are read-only. - Delete — removes the note permanently. The assistant stops seeing it from its next conversation onward.
The per-row reload icon appears on this grid but has no effect for memory notes — reload re-syncs live entities like devices and zones, and a memory note is just stored text. There is nothing to reload.
How notes reach the assistant
- Site notes are auto-loaded — newest 50. At the start of every conversation, the assistant pulls in the 50 most recent site-scoped notes (by Learned date) and treats them as known facts about this install. If you have more than 50 site notes, the older ones aren't in the always-on context — the assistant can still find them, but only by searching its memory when something seems relevant. Keeping the catalog lean (delete stale notes) keeps the most important facts in that always-on set.
- User notes are never auto-loaded. A
user-scoped note stays private to the person who taught it and is only surfaced when the assistant searches its memory for that same user. Usesitescope for anything everyone's assistant sessions should know; useuserscope for one person's personal shorthand. - Searching fills the gaps. Beyond the auto-loaded set, the assistant can search all stored notes (by keyword, by related entity, or to include the current user's personal notes) when it needs deeper context — that search is what stamps Last Recalled.
Related Documentation
- AI Assistant — setup and the full skill catalog for Ask GEM, including how it records and recalls these notes
- UIs — Ask GEM configuration per UI (hide the button, room-name resolution)