continuous-learning-v2

Capture Claude Code session behavior by enabling PreToolUse and PostToolUse hooks in your settings, which deterministically record tool calls and user corrections into local observations for learning reusable, confidence-scored atomic instincts.

Confidence-scored atomic instincts are small, learned behaviors converted from repeated session patterns, tracked with evidence trails and confidence scores to evaluate their reliability before clustering into higher-level skills or agents.

Continuous learning isolates project habits by storing instincts under a detected git project hash, ensuring safe cross-project isolation while allowing promotion to global scope only when patterns repeat reliably across multiple projects.

Review learned instincts by running the /instinct-status command, which displays project-scoped and global instincts, while the evolution pipeline supports reviewing, exporting, and importing the accumulated instinct library.

A git-backed project is required for project detection via environment variables and git metadata, enabling deterministic hook capture and safe YAML-frontmatter instinct persistence scoped to that specific project hash.

Background analysis can be run on locally stored session observations, processing scrubbed hook captures asynchronously to convert repeated tool use patterns into confidence-scored atomic instincts without blocking active sessions.