LoreTokens are an AI-native format that lets models work on compact meaning instead of raw data. One line encodes what pages of JSON cannot.
Instead of squeezing bytes, LoreTokens compress meaning. Regular, symbolic, and explicit -- the way language models already process information.
A single line encodes domain, concept, fields, and output shape. Modern LLMs expand it back into full detail on demand.
Atlas uses LoreTokens for exact recall across entire codebases. Nothing is approximated, nothing degrades at scale.
No decompression step. The compressed form IS the usable form. Models read LoreTokens directly -- no parsing overhead, no information loss.
Choose the density that fits the use case. All three tiers are interchangeable and convert losslessly between each other.
Human-friendly. For documentation, audits, and exported knowledge where clarity matters more than density.
Balanced. For CLI tools, diffs, repos, and day-to-day use. Compact but still readable.
Machine-centric. Maximum compression for hot paths, GPU hooks, and high-throughput pipelines.
Proprietary format. The LoreToken specification, glyph tables, GPU hook, and internal tooling are trade secrets. This page describes what LoreTokens achieve, not how they work internally.
Every product in the stack runs on LoreTokens.
LoreTokens power the entire ShipItClean adversarial code review platform and the SAIQL database engine.