github.devotd.
NewCoding agents

graphicle-ai/graphicle v0.1.1 — Rust Code Graph and Agent Memory

graphicle-ai/graphicle is a Rust single-binary tool that provides code graphs and memory to coding agents.

Repositorygraphicle-ai/graphicle ↗

Code graph and memory for your coding agents, in one Rust program. For Claude Code, Codex, Cursor and any MCP client.

stars
★ 261
Forks
15
Language
Rust
License
Apache-2.0
Latest release
v0.1.1
Read original source ↗View on GitHub ↗

graphicle-ai/graphicle bundles a code graph and memory for coding agents into a single Rust binary. It indexes a repository to build a code graph and collects documents, decisions, and past sessions in SQLite. It serves sourced excerpts to Claude Code, Codex, Cursor, and any MCP client, running locally without API keys or accounts. The latest release is v0.1.1. The provided information contains no change log for this version.

According to the README, it is a project that ports the Python code graph engine of Graphify 0.9.88 to Rust. It is closer to a language port of an existing engine than to new algorithms. The repository proposes a verification method that runs the upstream project's own test files and records whether they pass and, for files it does not run, the reason. If an answer's citations differ from the source, it rejects them; when only whitespace and indentation differ, it reverts to the original text. It also includes a design that links memory to code symbols and invalidates the link when a symbol changes. It is licensed under Apache-2.0 and its primary language is Rust.

In the README's own measurements, it reports being 3.0–22.7 times faster than Graphify 0.9.91 across 39 measurements over 13 codebases, and finding 99.2% true positives in a 30-question test. The measurement method and limitations are published in docs/benchmarks.md. However, it ties with grep on the code-finding test, and it does not validate long conversations or accumulated memory. The published figures should therefore be read within the scope of that test, while performance over longer usage and memory quality remain unverified.