Memory
Palui uses recall as long-term memory, with memory tools, optional auto-recall before a turn and auto-save at compaction.
Palui's long-term memory is recall, connected as a memory connector. Within a chat, the session (a summary plus the last 30 messages) is short-term memory. recall is what carries facts across sessions and chats.
There are three ways memory gets used.
Memory tools
The recall connector mounts recall's own MCP endpoint, so the model can call add_memory, search_memories, ask_memory, get_profile and list_memories when you ask it to remember or look something up. Forgetting a memory is a separate tool classed destructive, so it always asks first (see Approvals).
Auto-recall
When the Auto-recall switch is on, Palui asks recall for your profile, matched against your message, before each turn. The result goes into the turn's context.
- Timeout: 1.5 s. If recall doesn't answer in time, the turn goes on without memory.
- Skipped for commands and for messages under 4 words.
- The recall card shows the skip rate, for example "skipped 4 of last 20".
Auto-save
When the Auto-save switch is on, Palui saves to recall at session compaction (past about 40 messages) and when you send /new.
What gets saved: your own messages and the final replies.
What's never saved:
- Tool output. Results from MCP servers, web fetch or any other tool never go into recall. That keeps a poisoned tool result from becoming a permanent memory that auto-recall would feed back into every later turn.
- Tainted turns. A turn that read untrusted content (web fetch, a webhook payload, an
externalMCP server) is skipped entirely. See The taint rule.
Each save carries a source id built from the session and the message range, so saving the same slice twice doesn't create duplicates.
Turning it off
Each switch is separate: you can keep the memory tools and turn off auto-recall, or the reverse. Pausing the recall connector removes all three from the next turn on. See Saklar.