Skip to main content
Learnings is the company’s living memory of what works. Every time an Action is assessed with a conclusive verdict, Cortex distills a lesson in the format:
For problem X, doing Y → fixed it / made it worse / changed nothing.
That way, the next time the same problem shows up, no one has to “reinvent the wheel” or repeat a mistake already made. The agents also consult this base to act better.
Learnings base

📸 Screenshot to insert [/en/cortex/learnings]: Learnings page with a list of 'Problem → Action → Outcome' cards, the semantic search bar at the top and the button to refresh the base.

​
How it works, in one sentence

You assess actions in Actions → each conclusive verdict (positive/negative/neutral) becomes a learning problem → action → outcome → the base becomes searchable and feeds future decisions (yours and the agents’).

​
Anatomy of a learning

The concrete situation the action tried to solve, written in a generic, reusable way — no proper names or absolute numbers. E.g. “message delivery rate below target”, “infrastructure cost over budget”.
What was done, summarized in a canonical way. E.g. “migrate the workload to spot instances”, “switch the SMS provider”.
The result observed in the data: fixed it (the action improved the KPI), made it worse or changed nothing. It comes from the impact analysis verdict.

​
Semantic search (no need to hit the exact word)

The search bar understands meaning, not just the exact word. Describe the problem your own way — “delivery rate dropping”, “infra cost high” — and Cortex finds the related learnings, even if they’re written with other words.
If you don’t find it right away, describe the problem another way. The search is by meaning; rephrasing helps more than insisting on the same wording.

​
Conflicting outcomes

Sometimes the same action had opposite outcomes for the same problem at different times (it worked once, didn’t work another time). Cortex flags this conflict rather than hiding it — it’s a valuable signal: that solution depends on context, and it’s worth investigating when it works.

​
Refresh the base from history

The refresh (backfill) button sweeps the already-assessed actions and distills learnings that weren’t in the base yet. Handy right after you assess a large batch of actions, or to “pull” lessons from old actions.

​
Manage learnings

You can archive a learning that no longer makes sense or delete it (a permanent action). The base is always the company’s — one company’s learnings never leak into another.
“Too early” and “cannot assess” verdicts don’t generate a learning — the signal is still weak. Only conclusive outcomes enter the base, to keep it trustworthy.