Smriti turns reasoning into checkpoints

This feels like a legitimate missing primitive for anyone switching models mid-thought. The product will live or die on whether checkpointing feels lighter than the docs-and-copy-paste workflow it's trying to replace.

• –Manual checkpointing is the right bet; the README's abandoned auto-checkpointing experiment suggests the author is optimizing for signal, not spam.
• –The mounted/forked model is the interesting part: it preserves a clean branch, blocks future-session leakage, and makes replay/debugging far more realistic.
• –This is most compelling for agents and multi-model comparison, not as a generic memory layer.
• –The current scope is intentionally narrow: single-user, no merging, no streaming, and no integrations yet.