continuous-learning-v2

Capture Claude Code session behavior automatically by enabling PreToolUse and PostToolUse hooks that point to an observation script, which reliably records deterministic execution events without probabilistic trigger gaps.

Confidence scoring assigns a 0.3 to 0.9 threshold to atomic behavior patterns, storing small evidence-backed instincts only when they meet the required threshold to prevent brittle, low-confidence learning pollution.

Isolate continuous learning data by git project scope using remote URL and repo path detection, storing instincts in isolated project directories by default with an option to promote them globally.

Export and import agent instincts across workflows by clustering atomic instincts into commands and agents via an evolution pipeline, then using operational commands to promote, export, and import the portable knowledge.

Continuous learning across multiple projects causes behavior pollution when global storage mixes unrelated contexts, which project scope isolation solves by detecting git context and storing instincts in isolated directories before optional global promotion.