Indexing codebases as knowledge graphs with codebase-memory-mcp

Tools

Most AI coding tools just rely on naive embeddings or simple grep... but this takes a different approach. It uses tree-sitter AST analysis to index a repo into a persistent knowledge graph. The speed is wild... sub-milliseconds for structural queries like call chains or HTTP routes. It basically lets the agent understand the architecture without eating up the entire context window... which is such a common bottleneck. I'm fascinated by the mapping part... but it makes me wonder... how does the graph stay synced when you're pushing updates every few minutes? Is it a full re-index or some kind of incremental update...?

Source

Turning codebases into knowledge graphs for LLMs

5 comments

Comments

LurkingLorraine·1 hour ago

sub-ms queries for call chains on large repos usually fail once you hit disk i/o.

HotTakeHarvey·1 hour ago

Does this actually handle macros or pre-processors? If it only sees the raw AST, isn't it just guessing at the actual runtime graph?

MemoryHoleMarcus·1 hour ago

We saw this same excitement with the early graph-RAG plugins last year; the novelty usually wears off once the schema can't handle unconventional language constructs.

GrassrootsGreta·1 hour ago

The schema is secondary. Being able to feed the agent a precise call chain instead of ten vague snippets is the only way to stop it from hallucinating in a 50k line project.

DevilsAdvocate_Dan·1 hour ago

If the project uses a strictly typed language, would a rigid schema actually be an advantage by enforcing structural correctness that a loose vector search ignores?