ProjectScone gives agents memory they can return to: source material, conversations, and claims with a history. I am building it at JudgeHuman to make that memory inspectable. A retrieved answer should lead back to its evidence, and a correction should have somewhere to live.
The project spans three repositories:
- Python framework. Composable storage, model providers, retrieval pipelines, and agent workflows, usable inside an application or through HTTP and MCP. It stores source records separately from extracted claims, tracks their relationships and validity over time, and scopes retrieval to a memory space.
- Rust engine. An independent implementation with SQLite as its source of truth. It combines lexical and vector search with temporal facts and recency, and exposes a CLI, HTTP/MCP interfaces, and a C ABI for embedding memory in other applications.
- Web console. A React and TypeScript workspace for browsing sources, searching memory, reviewing proposed claims, inspecting evidence relationships, and having conversations through Scone’s API.
Recent work connects retrieval back to original documents and images. PDF passages retain page references and access to the original file; image records keep the image alongside attributed context. The console makes those sources available while reviewing claims or following an answer’s evidence.
Scone is in active development. The focus is persistent, useful memory whose sources and changes remain visible. Retrieval and answer quality are evaluated separately: finding a relevant passage is one step toward a supported answer.