I don’t do “trust me.” I do receipts. That’s why I built RFT Memory Receipt Engine — a local agent memory layer that generates a verifiable receipt on every assistant turn, showing exactly which memory slices were retrieved and proving they exist in a tamper-evident, append-only ledger.
What my space demonstrates: -
Durable session memory via append-only event logs - Fast lexical retrieval using SQLite FTS (no embeddings required) - Hash-chained ledger so edits/reordering don’t silently pass - Receipts per turn containing: retrieved event IDs/snippets + digests + chain hashes + prompt hash + response hash - Receipt verification: upload a receipt back into the Space and it confirms the hashes match real stored events Why this matters: most “memory” systems store chat logs and call it solved. The real problem is provable influence — when someone asks “what did the agent actually use to decide that?”, you need evidence, not vibes. Receipts make agent memory inspectable, auditable, and defensible.
Quick test prompts (copy/paste): -
Store this: Dog=Nova, City=Manchester, Drink=Pepsi Max. - What’s my dog’s name? - What city did I say? - My drink is Coke Zero now. This overrides earlier. - What’s my favourite drink?
Then download the receipt JSON and upload it into Verify Receipt. It should pass. Edit one character in the receipt and it should fail.
I encourage any feedback around What extra fields would make receipts more useful for audits/debugging? - Should I add optional semantic retrieval (embeddings) as a toggle, or keep it lexical-only for maximum reliability? - Any agent framework integrations you’d want next (LangGraph / CrewAI / custom?
Liam-RFTsystems