Learning Loop settings
Settings reference for the in-development Learning Loop: the master switch, which models review finished work and draft skills, the daily spending limit, what counts as a big task, how long a session must be quiet before it is reviewed, the daily new-skill limit, how much the chat says about what was learned, whether your scheduled and recipe runs are also reviewed, how long the history is kept, how big a memory area may grow before near-duplicates are merged, whether project skills are shared with other coding engines, and whether the reviewer sees tool activity. Off by default.
On this page
- Learning Loop
- Review model
- Skill drafting model
- Daily spending limit
- Big task: tool steps
- Big task: minutes
- Quiet time before review
- New skills per day
- Chat notes
- Also learn from my scheduled and recipe sessions
- History kept
- Memory size limit
- Merge near-duplicate memories
- Share project skills with other coding engines
- Let the review see tool activity
The Learning Loop lets agents look back over finished work in the background. Hard tasks can become reusable skills, and facts about you can become memories. Every change is checked, needs your approval where it matters, and can be undone. The feature is still in development, so it stays off and hidden until it ships.
Learning Loop
The master switch. When on, Omniscio reviews finished sessions in the background and proposes skills and memories from them. When off, nothing is reviewed and nothing is spent.
Accepted values: on / off. Off by default.
Review model
The model that reads a finished session and decides whether anything is worth learning. "Utility" follows the global Utility AI model you chose in AI settings, so it works with whichever account you have. You can also pin Haiku, Sonnet, Opus or Luna.
Accepted values: utility, haiku, sonnet, opus, luna. Utility by default.
Skill drafting model
The model that writes a new skill once a review decides one is worth having. "Best available" picks the strongest model your accounts can reach. You can also pin Haiku, Sonnet, Opus or Luna.
Accepted values: best-available, haiku, sonnet, opus, luna. Best available by default.
Daily spending limit
The most the Learning Loop may spend in one day, in US dollars, across reviews, skill drafts and skill checks. When the limit is reached it pauses until the next day.
Accepted values: 0.1 to 100. 1 by default.
Big task: tool steps
How many tool steps (file edits, commands, searches) a session needs before it counts as a big task worth learning from.
Accepted values: whole numbers from 5 to 500. 20 by default.
Big task: minutes
How many minutes of work a session needs before it counts as a big task worth learning from.
Accepted values: whole numbers from 1 to 480. 15 by default.
Quiet time before review
How many minutes a session must sit idle before it is reviewed, so a session you are still working in is never reviewed halfway through.
Accepted values: whole numbers from 5 to 1440. 30 by default.
New skills per day
The most new skills the Learning Loop may draft in one day. Set it to 0 to stop new skills while still allowing memories.
Accepted values: whole numbers from 0 to 50. 5 by default.
Chat notes
How much the chat says when something is learned. Off shows nothing, On adds one short line, and Detailed adds what was learned and why.
Accepted values: off, on, detailed. On by default.
Also learn from my scheduled and recipe sessions
Lets the Learning Loop also review the runs of your own scheduled jobs and recipes, which nobody was watching. Skills only, never memories. Every skill these sessions suggest waits for your approval, and each scheduled job or recipe is reviewed at most once a day, after it finishes. Recipes started by email, messages, boards, other agents or automations are never reviewed.
Accepted values: on / off. Off by default.
History kept
How many days the record of what was learned, approved and undone is kept before it is cleared. The newest entry of every learned skill that is still live is always kept, so you can still view, undo or edit it.
Accepted values: whole numbers from 7 to 3650. 90 by default.
Memory size limit
How much saved memory one area (about you, your people, or a single project) may hold before the Learning Loop merges near-duplicate memories into shorter ones. Nothing is merged while an area is under this size. It is counted in characters of memory text.
Accepted values: whole numbers from 5,000 to 200,000. 30,000 by default.
Merge near-duplicate memories
When an area is over its size limit, a background run merges memories that say nearly the same thing into one, in up to three rounds. Every merge is a line on the Learning timeline ("Memories combined — N into 1") that you can Undo, which brings back every original; Restore redoes it. A merge you undid is never redone. It uses a little of the daily spending limit. With no AI key, or the limit spent, it writes one "Couldn't combine" line per area per day and merges nothing.
Accepted values: on / off. On by default (once the Learning Loop is on).
Share project skills with other coding engines
Lets Codex, Gemini and Hermes sessions use a project's own skills and the skills learned for it, not
just Claude sessions. It works only while the Learning Loop switch in Settings → Lab is on and
Sync skills under Settings → Automations & Rules → Sync to Other AI Tools is on too. Codex and Gemini get copies in the project's
.agents/skills folder, kept out of git; Hermes reads the folders in place. Your global skills folders
are never touched.
Accepted values: on / off. On by default.
Let the review see tool activity
Shows the reviewer which tools a Claude session ran and what came back, with secrets removed, so it can learn from what actually happened. Each call is cut short (input at 600 characters, result at 1,000) and the reviewer sees about 12,000 characters of them at most, failed calls first. Tool lines never count as evidence for a memory. It does nothing for other engines' sessions.
Accepted values: on / off. On by default.