Tune memory recall, saving and cleanup
These instance-wide limits bound what memory costs and how much it can change, while each account keeps its own switches.
Turn-start recall
Settings → <assistant name> AI → Limits holds Turn-start memory wait: how long the first response of a turn waits for the opening memory search, from 0 to 30 seconds, 5 by default.
When the search is slower, the reply starts without it and the results arrive at the next model call. Set the value to 0 and the first response never waits.
Recall while Chetty works
Setting (Settings → <assistant name> AI → Limits) |
Default | Range |
|---|---|---|
| Recall while working — searches | 10 | 0–20 |
| Recall while working — memories per batch | 2 | 0–10 |
| Recall while working — turn budget | 20,000 UTF-8 bytes | 10,000–40,000 |
Setting searches to 0 switches recall while working off, and so does 0 memories per batch. Searches also caps the embedding requests one tool loop may make, and a new steering message gives the turn a fresh budget.
Auto-save writes
Settings → <assistant name> AI → Runtime holds Maximum memory writes per exchange, 0 to 6, 2 by default. It limits how many memories automatic saving may add, revise, merge or delete after one exchange; 0 switches auto-save off entirely and leaves the memory tools working.
Relevance and duplicates
Three similarity thresholds decide which memory counts and which is redundant, under Settings → <assistant name> AI → Runtime:
- Memory relevance floor — the minimum similarity at which a memory counts as related. Anything below is left out of recall and of the memory search box.
- Duplicate threshold when saving — at this similarity, automatic saving treats a new memory as a rewording of the stored one. Default 0.93.
- Duplicate threshold when recalling — at this similarity, an already chosen memory makes another one redundant, so only one takes a recall slot. Default 0.70.
Recheck the saving threshold after changing the embedding model or after memory has grown substantially.
Maintenance thresholds
Automatic memory maintenance per account waits for two conditions, both under Settings → <assistant name> AI → Runtime:
- Memory dreaming: hours since the last pass — 1 to 168 hours, 24 by default.
- Memory dreaming: conversations since the last pass — 1 to 50, 5 by default. One busy conversation is not yet a sign that stored memories contradict each other.
Both are read at every check, so a change applies to the next pass without a restart. Run now in the account ignores both. See Skill learning and memory maintenance.
Retention
Open Memory retention and click Edit retention for the automatic cleanup:
| Setting | Default | Range |
|---|---|---|
| Memory retention enabled | on | on or off |
| Grace period | 14 days | 0–365 |
| Vitality floor | 10 | 1–90 |
| Daily trash limit | 5 | 1–200 |
| Time to halve, importance 1 to 10 | 1, 3, 7, 14, 21, 30, 45, 60, 90, 180 days | 0.1–365 each |
The master switch pauses all automatic removal, whatever the other values say. Neither the floor nor a half-life accepts zero. Read When memories fade and how to restore them for what the numbers do to an account's memories.
Skill review
With the Skills plugin enabled, learning from finished work is bounded by three values under Settings → <assistant name> AI → Runtime:
- Minimum tool calls for a skill review — 1 to 200, 1 by default. A shorter exchange is neither reviewed nor counted.
- Skill review every N tool calls — 1 to 200, 100 by default. Counted across exchanges; an interval an account saved for itself stays unchanged.
- Maximum skill changes per exchange — 0 to 6, 1 by default. 0 switches learning off for everyone; the manual skill tools keep working.