Skip to main content
ngram’s learning mechanisms operate on runtime state and recorded experience. They do not train or fine-tune the inference model’s weights.

​
Turn outcomes

The outcome store records fields such as route, tools used, tool success rate, user signal, outcome score, and an optional embedding. Queries can retrieve recent or similar outcomes and aggregate tool effectiveness. The score is an internal heuristic. It can help compare behavior within an evaluation, but it is not a calibrated measure of task correctness or user satisfaction.

​
Correction-driven learning

The runtime can classify a later user message as corrective and retain useful content as a belief or procedural skill. This makes a correction available to future retrieval. Retention does not guarantee that the same mistake will never recur. Classification, extraction, retrieval, and generation can each fail. Review the stored artifact and test a follow-up situation when the correction matters.

​
Tool experience

The self-model records tool usage and failures and can supply hints about reliability in later turn context. This gives the model evidence about repeated failures and successful tools. Those hints should be read as observations about the configured environment. A tool failing on one backend does not prove the interface is intrinsically unreliable. Procedural memory stores reusable guidance. Distillation extracts material from experience. Trait evolution adjusts identity-related runtime state. Each has a different role and operating cadence. For useful evaluation, record the task, configuration, retained artifacts, and resulting behavior. Change one mechanism at a time and keep a reproducible example of both the original failure and the expected improvement. Source: memory/turn_outcomes.py and entity.py.