Researchers from Hugging Face have unveiled CodeNib, a novel multi-view data system designed to streamline how coding agents access and maintain repository context. The system addresses a fundamental inefficiency: agents repeatedly search, navigate, and retain context from evolving codebases, often relying on disconnected indexes and task-local histories that force redundant discovery.
CodeNib treats repository context as a data-systems problem. It builds three types of views—lexical, dense (vector), and structural—for each commit. These views are maintained incrementally and served through a single runtime that supports ranked search, symbol navigation, and bounded context retrieval. The system maps all outputs to repository-relative source ranges and maintains selected views across edits.
In evaluations across 100 repository snapshots, CodeNib demonstrated significant performance gains. When outputs matched an independent rebuild, graph updates were 8.7× faster and vector updates 25.4× faster at the median. On static-navigation requests that matched normalized live-server locations (63% of 1,000 requests), the median per-request live-to-static latency ratio was 4.7× in favor of CodeNib.
Further testing across five agent models showed that CodeNib's context policies preserved localization quality while using 50–87% fewer trajectory tokens compared to paired grep/read approaches. The system establishes explicit, operation-specific validity boundaries for each view, ensuring consistency and reliability.
The code and artifacts are available on GitHub at github.com/sysevol-ai/CodeNib. The paper also highlights related works including Agent Retrieval Bench and Agent Memory, indicating growing interest in optimizing agent-context systems.