Codebase-Memory-MCP review: setup and trade-offs

Codebase-Memory-MCP gives a coding assistant a persistent graph of code structure to query across sessions. The speaker describes an MIT-licensed C binary with zero runtime dependencies, local processing and SQLite storage. It includes 158 Tree-sitter grammars, with hybrid LSP type resolution for a dozen major languages.

The walkthrough uses Django. The presenter reports indexing 4,787 files into 55,456 nodes and 343,912 edges in under a minute in their sandbox; the vendor reports about six seconds on an Apple M3 Pro. Examples cover caller tracing, read-only Cypher-style queries, architecture summaries and change-impact analysis. A compressed graph snapshot lets teammates start from an existing index.

The performance figures need context. The project's five-question benchmark reports about 3,400 tokens versus 412,000 for a naive file-by-file baseline. The cited preprint reports roughly tenfold token savings across 31 repositories, but answer quality falls from 92% with file exploration to 83% with graph queries. A focused file reader could narrow the savings.

Semantic scoring requires moderate or full indexing; the presenter reports irrelevant results in default mode. Cypher support is a limited read-only subset. The review also distinguishes the graph from Claude Code's prose instructions and memory notes: it complements those features. The speaker sees the strongest fit in large repositories with repeated agent sessions.