Spaces & memory
Spaces layer shared context over your chats. Memory is learned from completed replies - per Space (on by default) and globally (off until you turn it on) - and has hard caps.
Memories aren’t being saved
You expect memory to pick up facts, but nothing appears in the Space’s Memory tab or the global memories list.
- Memory extraction only runs after a complete assistant reply. Cancelled or errored turns are skipped. It also waits until the model is free - no other reply running or queued - so facts can turn up a little later; turns still waiting are kept if you quit, and picked up on the next launch.
- In a Space, Memory must be on in the Space’s Settings tab. Outside a Space, nothing is saved until you turn on Settings → Features → Global memories. Space chats never write to global memories, and Private Chat never saves anything.
- Only what you typed counts. Facts in attached files or fetched web pages are ignored, and so are requests (“write me a script”) - most turns produce nothing, and that’s normal. To save something specific, use
/remember <fact>. - Passwords, PINs, API keys and other credentials are never saved, even ones you share about yourself - memory is stored as plain text and sent to models. Facts you add yourself aren’t filtered.
- A fact that repeats one you already have is dropped, and each fact the extractor saves must be at most 280 characters.
- Memory uses the same provider and model the chat is using, and gives up after a minute. Tiny or slow models often return empty extractions; try a larger one.
Memory captured something wrong or private
A bad fact landed in long-term memory.
Right after it lands, click Undo on the Saved to memory toast. Later, open the Space’s Memory tab - or, for global memories, Settings → Features → Manage global memories - and click the fact’s text to edit it, or hover the row and click the trash icon to delete it. /forget <text> works too. If memory removed or rewrote a fact it shouldn’t have, use Undo on the Removed memory or Updated memory toast.
Turning a Space’s memory switch off only stops new facts; the ones already saved are still sent until you delete them. Turning off Global memories stops both saving and sending.
“Memory is off” when using /remember
/remember answers Memory is off for “…” or Global memory is off and saves nothing.
/remember follows the same switches as automatic memory. In a Space chat, turn on Memory in the Space’s Settings tab (the Change in Settings link on its Memory tab goes there). Outside a Space, turn on Settings → Features → Global memories, or run the command in a chat inside a Space. The Add field in the memory editor is unavailable for the same reason.
A memory is marked “not sent”
The memory editor tags some older facts not sent, and the model doesn’t seem to know them.
Each memory list sends only its newest facts, up to 60 facts and about 6,000 characters, so it can’t crowd a small context window. Older facts are kept but not sent. Delete facts you no longer need (or shorten long ones you typed yourself) and the older ones move back into the prompt.
Reference sources won’t add to a Space
Adding a source to a Space fails or silently does nothing.
Each upload can be at most 20 MB (… exceeds the 20 MB limit), and a Space’s uploads at most 200 MB in total (… would exceed the 200 MB space limit; at the limit, the Sources tab reads Reached the 200 MB upload limit and Upload is disabled). Library items linked with From Library don’t count toward that limit. Remove older uploads, or move the content into a smaller text file - or into your Library, and link it from there.
Each file is read on its own, so one that fails doesn’t stop the rest, and the error lists each failed file on its own line. A scanned PDF, an empty Word document or an empty text file is refused with … no readable text (- it may be a scanned PDF for a scan): add its pages as images instead, or a text version.
If a source was added but the model ignores it: a source switched off on the Sources tab isn’t sent, and formats Loach can’t read (spreadsheets, archives, legacy .doc files) are kept but marked unreadable and never sent - see A Space upload says “unreadable”. A long source may reach the model only as excerpts - see The model ignored my source.
Space instructions seem to override my custom instructions
Your global Custom Instructions, or a chat’s own instructions, don’t appear to apply in chats inside a Space.
This is by design. When a Space has its own instructions, they replace both the global ones and the chat’s own Additional instructions for chats in that Space. Either clear the Space’s instructions, or repeat the relevant parts in them. See Custom instructions: precedence.