What problem does it solve? AI agent configurations accumulate token waste: heavy CLAUDE.md files, never-invoked skills, and context re-pasted every session. The caveman learn command measures where tokens go, but acting on its findings requires careful, reversible edits that actually lower cost without degrading the agent's recall. ## Core Features & Use Cases - Consent-gated fix loop: Reviews ranked token sinks from caveman learn report --json and proposes one fix at a time, applying each edit only after the user says yes. - Reducible trims with a net-token-negative gate: Materializes a candidate via caveman learn apply <sink_id> --dry-run, applies the trim, then re-measures and reverts if tokens per turn did not drop. - Recurring-context offload to cavemem: Verifies the block's sha256 against the locator, stores it with caveman mem remember, writes a recall pointer, and reverts entirely if recall fails so the agent is never made dumber. - Skill distillation holdouts: For repeated tool-step procedures, writes a skill and starts an on/off experiment with caveman learn experiment, requiring at least 5 sessions per arm before any verdict. - Honest savings reporting: Presents caveman learn savings grouped by measurement rung (deterministic remeasure, controlled holdout, counterfactual replay, interrupted time series) without summing across rungs or claiming verified figures. - Use Case: After a learn report flags a 400-line CLAUDE.md block re-established every session, offload it to cavemem, leave a pointer, confirm recall works, and record the applied fix for longitudinal verdicts. ## Quick Start Ask the agent to run a caveman learn report and walk you through fixing the top token sinks one approved edit at a time.
