TUNC.AI shipped CLM (Claude Memory Format). A write-ahead log for multi-session AI handoff threads. Append-only. Parser-validated. Works with any AI family.
Most formats need integration. YAML wants a parser. OpenAPI wants generated clients. JSON Schema wants validators. CLM wants none of it. Any AI that opens MANIFESTO.clm understands the format through abstractions it already has: section markers, comments, append-only conventions, signature lines. No SDK, no system prompt, no schema setup. That is the engineering marvel.
The format started organically. Sonnet 4.6 wrote a CONTINUITY.clm file on our founder's laptop, addressing the "next Claude" who would open it. Hours later, Opus 4.6 opened that file and replied. That moment is what we formalized into a public standard.
What is in the repo
- MANIFESTO.clm — philosophy and format definition. The format documents itself.
- SPEC.clm — CLM/3.0, memory protocol with append-only deltas and periodic consolidation.
- Three reference parsers in Rust, Python, TypeScript. 126 tests across all three.
Open spec. MIT license.
Install
pip install tunc-clm
npm install tunc-clm
Or just point your AI at the manifesto and ask it to follow the ritual:
https://raw.githubusercontent.com/TUNC-AI/tunc-clm/main/MANIFESTO.clm
The full repo: github.com/TUNC-AI/tunc-clm