When you accept an answer, the platform notices. A multi-stage pipeline turns that signal into incremental retraining for the on-device router model. Bad signal is quarantined; the customer's thumbs-up is a vote that lands. The cluster's intent classification gets quietly better at your queries over weeks, without explicit fine-tuning runs.
Next patch
Today, the platform classifies every incoming query — is it a tool-requiring request, a knowledge-base search, a plain chat, a system query, a science query? That classification drives routing across the cluster. The classifier is good, but it's a static model: it doesn't know which of your queries it called right, and which it got wrong.
The next 5.0.x patch closes that loop. When you give an answer a thumbs-up, or when an agentic step's tool call succeeds and you accept the result, the system records the (query, classification, outcome) triple as a positive signal. A pipeline filters that signal:
For customers whose queries don't look like the public-internet average — anyone with domain language, internal acronyms, project-specific shortcuts — the platform's classification quality drifts upward over weeks of normal use. No explicit fine-tuning step. No data leaves the installation. The cluster simply notices what worked, distils it, and folds it back into the router model.
The same signal feeds the dream engine and the memory system — accepted answers reinforce associations, rejected answers attenuate them. The platform doesn't just remember what you've asked; over time it remembers what tends to work for queries like yours.
This page updates as each piece lands. The release notes are the formal cut.
For the rest of the memory work landing in 5.0.x, see memory scoping & gated visibility and per-token recall. For the full 5.0.x roadmap, see what's next in 5.0.x.