
2022 · 202617 stars
Corollary
An agent runtime that keeps every conclusion tied to the evidence behind it.
Agents usually keep their state as a message log, so a wrong fact at step 3 is read at step 40 with the same trust as a verified tool result. Corollary stores beliefs instead: every claim records what it follows from, and a truth maintenance system retracts each conclusion that depended on a corrected input and re-derives only those. The model just proposes claims; the runtime validates them, runs the tools and builds every context from the beliefs still in force.
- Correct one input and only its dependents change: 12 of 20 conclusions in the demo, with zero model calls
- Answers ship with a proof a deterministic verifier checks: arithmetic, citations and dates
- Python
- mypy strict
- Claude
- OpenAI-compatible
- Ollama
- MkDocs















