CodeMemory 0.6.0

CodeMemory

Local-first repository intelligence engine exposed via the Model Context Protocol (MCP). Build a persistent semantic memory layer over any codebase.

Build architecture-aware AI agents that understand symbols, dependencies, semantics, and git history — without coupling to a specific database, LLM, or embedding provider.

Installation

dotnet add package CodeMemory

What You Get

• Semantic code search — natural language queries over indexed code via IEmbeddingGenerator<string, Embedding<float>>
• Dependency graph — upstream/downstream symbol relationships (calls, references, inheritance, implementation)
• Architecture overview — component structure, language breakdown, file and symbol counts
• Impact analysis — change impact with downstream dependencies, affected files, components, and test coverage
• Component clustering — logical groupings based on inter-component coupling density
• Symbol history — per-symbol git commit history with authors and timestamps
• Hotspot detection — most frequently changed files ranked by commit count
• Edit context — comprehensive context for a symbol: source code, dependency chains, related symbols, and test coverage
• SQL queries — compose arbitrary filters across indexed fields using SQL (SELECT, WHERE, ORDER BY, GROUP BY, HAVING, CTEs, derived tables, aggregates, INNER/LEFT/RIGHT/FULL OUTER/CROSS JOINs, USING(col), UNION/INTERSECT/EXCEPT, IN (SELECT ...) subqueries, vector search via ORDER BY Similarity DESC) — parsed by SqlParserCS, executed as LINQ over InMemoryVectorStore
• Multi-repo support — per-repository storage isolation via ServiceRegistry + IRepoContextAccessor
• Pluggable storage — in-memory (InMemoryVectorStore, default, zero-dependency), SQLite (Microsoft.SemanticKernel.Connectors.SqliteVec), PostgreSQL with pgvector (Microsoft.SemanticKernel.Connectors.PgVector), or SQL Server (Microsoft.SemanticKernel.Connectors.SqlServer)
• File watcher — post-indexing FileWatcherService auto-reindexes changed/created/deleted files via FileSystemWatcher with debounce coalescing
• Multi-language parsing — Roslyn (C#) + Tree-sitter (TS/JS/Java/Python/Go/Rust/C/C++) for symbol extraction and relationships; .txt/.md indexed as text chunks

Dependency Injection

// Minimal setup — in-memory storage, offline n-gram embeddings
services.AddSingleton<IStorageService>(CreateInMemoryStorage(/* repoRoot */));
services.AddSingleton<IEmbeddingGenerator<string, Embedding<float>>, NgramEmbeddingGenerator>();
services.AddSingleton<DependencyGraphService>();
services.AddSingleton<ArchitectureService>();
services.AddSingleton<ComponentClusteringService>();
services.AddSingleton<GitHistoryService>();
services.AddSingleton<SemanticSearchService>();
services.AddSingleton<SqlQueryService>();
services.AddSingleton<CollectionRegistry>();

Storage is registered via factory methods (see CodeMemory.Storage.ServiceCollectionExtensions or CreateInMemoryStorage on IServiceProvider). For the ASP.NET multi-repo host, storage is registered at runtime via IServiceRegistry.Register(name, storage) and routed through StorageServiceRouter.

Key Dependencies

Targets

• Repository indexing, semantic analysis, and MCP tool exposure
• AI coding agents and IDE assistants needing architecture-aware context
• Multi-repo deployments with per-repository storage isolation

Learn More

• Full documentation and architecture: ARCHITECTURE.md
• AI agent engineering guidelines: AGENTS.md

License

MIT