Learning Behavioral Rhythms

Temporal Pattern Mining

Today’s major update closes the ambient inference feedback loop. Chalie no longer relies solely on config-based circadian heuristics to predict user energy or attention. The system now observes its own inferences—place, energy, mobility, and tempo—over time to mine statistical patterns. This allows it to learn that a specific user might have high energy at 10 PM despite general biological norms, or that they are usually in a specific (but privately hashed) location on Tuesday mornings.

I implemented a TemporalPatternService that uses Laplace-smoothed probabilities to handle low-data scenarios and established a privacy-first schema. All location data is HMAC-SHA256 hashed using a per-instance key before storage, and time is bucketed into local hours. These learned rhythms are now fed back into the AmbientInferenceService for better predictions and the CognitiveDriftEngine to seed more relevant background thoughts.

Tooling and Trust

The Google News tool has been renamed to news_tool and promoted to a “trusted” status. It is now installed by default for all users. This involved refactoring the manifest, removing old sandbox blocks that are no longer necessary for trusted tools, and setting up automated release workflows. The tool provides multi-source coverage without requiring an external API key, making it a reliable baseline for the research category.

Refinement and Onboarding

On the UI side, I pushed a small but necessary fix to the Brain interface. The sticky navbar was previously too transparent, causing it to blend into content during scrolls. Increasing the background opacity and adding a backdrop blur provides much-needed visual separation while maintaining the dark aesthetic.

Finally, I expanded the _ONBOARDING_SCHEDULE. The system was already capable of nudging for various user traits, but the schedule was missing entries for several core identity markers. Chalie will now slowly learn about a user’s occupation, communication preferences, and primary goals over the first few dozen interactions, allowing for a more tailored experience without an overwhelming initial setup.