My Agent Kept Forgetting Everything. So I Built a Sloth — a Zero-LLM Memory That Never Re-Explains Itself

Liquid syntax error: Unknown tag 'devchallenge'

Comments

[tercel]

That “85% coverage, 100% type hints” line quietly says a lot: the guarantees you describe (integrity, authorship, provenance) are being treated like a safety boundary, not a convenience feature. If 15% of the code isn’t under tests, a natural question is: is the untested surface only wiring/CLI, or are any verification or crypto paths in that slice?

The sha256:... commit IDs plus Ed25519 signatures and JCS imply a few nice properties right away:

• Anyone can recompute the hash and know if a memory commit changed at the byte level.
• Anyone with the public key can check that some holder of the private key signed that history.
• The parent hash link gives you a linear, append-only log as long as you never rewrite parents.

Things I’d be curious to poke at next:

• What’s the story for forks? If two processes commit with the same parent, do you treat that as a branch, an error, or something to reconcile?
• Can you cheaply verify “prefix integrity” (up to commit N) without walking the full chain every time, or do you lean on periodic checkpoints?
• For alethech search, how big can a history get before bounded lexical recall starts to drag, and do you expose any index stats to the agent?

Love that you leaned all the way into “bookkeeping, not cognition.” This feels closer to git + age for agents than yet another vector store with a marketing site.