Skip to main content

​
Learning

WednesdayAI can learn from your corrections and its own mistakes. Correct it a few times and it writes down a lesson. When a lesson keeps coming back, the AI drafts a proposed new skill - and nothing changes until you approve it. This is different from memory, which stores facts you ask the AI to remember. Learning is about behaviour: getting better at how it works for you.

​
What happens after you correct the AI

  1. When a chat finishes, the AI sets aside a short note of what happened - especially if you corrected something.
  2. If similar notes keep coming back, they are turned into a lesson. A lesson is one sentence the AI believes will help next time.
  3. Once a lesson has come back often enough (by default three times - only twice when the lesson started with you directly correcting the AI), the AI queues a skill proposal: a draft of a new reusable ability.
  4. Nothing is installed until you approve the proposal. The AI never writes a new skill on its own.
Learning may not be switched on for your AI - it starts switched off. If nothing seems to be learned, ask the person who runs your WednesdayAI to turn it on.

​
Reviewing what the AI learned

If you use the WednesdayAI control panel (the web page for your AI - ask the person who runs it for the link), open the Learning tab. There you can:
  • search the lessons the AI has written down,
  • open the chat a lesson came from,
  • preview a skill proposal in full before deciding, and
  • approve or reject several proposals at once.
You can also just ask in chat, for example:
“What have you learned recently?”
“Show me my learning proposals.”
“Apply proposal fix-python-quotes.”
The chat options only work for the owner of that WednesdayAI - the person it belongs to. In a group chat, other members cannot approve or change lessons.
1

Ask what is waiting for you

Send: “Show me my learning proposals.” You should get back a list of drafts, each with a short name and a summary of what it would do. If the list is empty, there is nothing waiting - that is normal after only a few chats.
2

Preview one proposal

Send: “Inspect proposal fix-python-quotes.” You should see the full text of the draft, plus a note if the built-in safety check flagged anything in it.
3

Decide

Send: “Apply proposal fix-python-quotes” to approve it, or “Reject proposal fix-python-quotes” to decline. You should get a confirmation either way.
Applying a proposal gives the AI a new skill that affects every future chat. Read the preview first. A rejected proposal stays on record - it is recorded, not erased - so you can always see what you declined.

​
If the safety check flags a proposal

Every proposal is scanned before it can be applied. If the scan finds something unsafe, the proposal is parked in quarantine instead. A quarantined proposal cannot be applied as-is - if you truly want it, ask the AI to create the proposal again after fixing what was flagged.

​
Fixing a lesson that is wrong

Lessons are never deleted. If the AI learned something wrong, tell it:
“That lesson about me liking plain text emails is wrong - I prefer HTML.”
The AI writes a replacement lesson and keeps the old one on record, so there is always a trail of what it believed and when. The outdated lesson stops being used.

​
Reading the full notebook

Lessons are also written out as readable files in your AI’s notes folder (inside what your operator calls the “workspace”). If you like reading files, ask the person who runs your WednesdayAI to show you the learning notes.

​
Privacy

Before anything is stored, the text passes through a redaction step that strips things like passwords, API keys, email addresses, and phone numbers. So if you accidentally paste a secret while correcting the AI, the stored lesson will not contain it. Still, avoid pasting secrets into chat in the first place.

​
If learning stops working

If the AI stopped picking up corrections after an update, the learning part probably failed to start with the update. Tell the person who runs your WednesdayAI to restart it and check that learning started without errors. Your existing lessons stay put either way.

​
What learning is not

  • It is not a place for secrets. Redaction helps, but do not rely on it.
  • It is not instant. Lessons need repetition by design - one accidental correction should not rewrite behaviour.
  • It is not memory. For facts you want remembered word-for-word, say “remember this” instead (see Memory).