Core concepts

Fidelity

Every number is labelled by how it was measured: exact (read from the tool's own usage records), approx (derived), or estimated (e.g. char-count fallback). You always know how much to trust a figure.

Identity

Usage attributes across four levels — org → account → system (machine) → person (os user) — so the same person using two AI accounts on two machines still rolls up correctly. Shared accounts (one login on many seats) are surfaced for license planning.

Sessions

A session is one continuous piece of AI work (a Claude Code session, a Codex thread, a chat conversation). The Brain captures sessions as turns (your steering + the model's responses), the tools and files they touched, and the artifacts they produced.

Trajectory — the build-story

Each session is distilled into its arc: specify → build → correct → verify → accept, with the corrections, pivots and verifications marked. It's how a piece of work actually unfolded.

Working-style

Across your turns, TokenTracker derives your steering fingerprint — how much you specify vs. steer vs. correct vs. verify — plus signature phrases and traits ("demands verification", "plan-then-delegate"). Stable across projects, re-derivable from raw.

Knowledge graph & bridges

Sessions link to projects, tools, models and files. A file touched in two sessions becomes a bridge — literally how your work carries forward across time and tools.