ProjectScone · A JudgeHuman project by Mark Sturman
The native Rust temporal memory engine, CLI, HTTP/MCP server, C ABI and benchmark harness. This Cargo workspace builds independently. The Python framework lives in the separate ProjectScone repository; the React frontend lives in ProjectScone-Webapp. Neither checkout is required for Rust builds.
Install a stable Rust toolchain with support for edition 2024 and a native C/C++ compiler, then run from this checkout:
cargo build --locked -p scone-cli --no-default-features
./target/debug/scone --embedder hash --llm none add --note "hello, memory"
./target/debug/scone --embedder hash --llm none search "hello"The deterministic hash embedder is useful for development and conformance; it
is not a semantic embedding model. These commands use the default local data
directory (~/.scone); choose --data-dir /path/to/local/data and --space NAME
to keep independent stores and scopes.
Default features add PDF text extraction and local ONNX embeddings:
cargo build --locked -p scone-cli
./target/debug/scone add --note "changed the oil on the truck"
./target/debug/scone search "vehicle maintenance"The local embedder downloads a model on first use, then reuses cached files. Default-feature compilation can also obtain ONNX runtime binaries. Use the minimal build above when model provisioning is unwanted. PDF text extraction is not OCR. Configure any LLM or external connector explicitly; basic ingestion and retrieval do not require an LLM service.
| Package | Purpose |
|---|---|
scone-core |
SQLite-backed episodes and temporal facts, indexing and recall |
scone-cli |
scone binary, HTTP server, MCP, hooks and embedded console |
scone-ffi |
C ABI subset; declarations in crates/scone-ffi/include/scone.h |
scone-bench |
Rust evaluation harness |
Recall combines lexical and vector results with facts and recency. Tags and
source filters narrow evidence. Recorded facts and model extractions remain
claims, not independently verified truth. --as-of queries validity intervals;
it does not reconstruct the database's past state before later corrections.
Returned/stored byte reduction is not measured token reduction or answer quality.
Rust and Python have distinct implementations and storage, with only a bounded
episode-transfer conformance profile rather than complete portability.
Run scone --help or scone <command> --help for supported commands. For an
explicitly model-free local console or MCP server:
./target/debug/scone --embedder hash --llm none ui
./target/debug/scone --embedder hash --llm none --space myproject mcpThe console binds to loopback and uses a process-scoped credential. Its native console and preview playground are embedded from existing HTML files at compile time. Building Rust does not rebuild the frontend. See CONTRIBUTING.md for the explicit asset-update procedure, native checks and external conformance. The separate Webapp's Python-oriented Memory page has not been established as a compatible replacement for the Rust native console.
cargo build --locked --release -p scone-ffi --no-default-featuresThis produces a native shared library in target/release (libscone_ffi.dylib
on macOS, libscone_ffi.so on Linux). The C ABI exposes a subset of the engine;
follow its header's ownership and error rules. A model-free FFI build requires
the hash embedder selection rather than local ONNX.
The release workflow packages the CLI, C header, license, citation and artifact notes for Apple Silicon macOS and x86-64 Linux. It does not package a compiled FFI library. Manual workflow runs produce downloadable CI artifacts; version tags publish releases. No frontend or Python build runs during packaging.
See CONTRIBUTING.md for setup and review requirements. The ProjectScone Research Attribution License is custom and MIT-derived, with mandatory research and academic citation. Credit Mark Sturman, JudgeHuman and ProjectScone. CITING.md provides MLA, APA, Chicago and BibTeX examples; CITATION.cff provides metadata.