Project
Engram
Local, persistent codebase memory any AI can query via MCP.
Engram is an open-source, local-first memory layer that indexes an entire codebase into a searchable vector store and exposes it to AI assistants, Claude Desktop, Cursor, any MCP client, over the Model Context Protocol. It's built for developers tired of re-explaining their architecture every session, and ships as a published Python CLI and MCP server (engram-dev on PyPI).
Stack
PythonMCPFAISSsentence-transformersPyTorchClick
What I built
05 systems, built solo, the parts that mattered.
AST-aware code chunking, not blind text splitting, it parses Python via the stdlib ast module to extract functions and classes with signatures, decorators, and docstrings, with regex extractors for JS/TS and a text fallback.
Four purpose-built MCP tools, query_memory, query_recent, whats_changed, and explain_file, going beyond a single generic search endpoint to give temporal and per-file capabilities.
Crash-proof incremental indexing, files process in batches of 100, persisting the FAISS index and hash registry after each batch to enable resume-after-interruption and change detection.
Git-aware temporal memory, parses recent commits, file history, and diff summaries to power what-changed-recently search, and still works without git.
Fully local, cross-platform, embeddings run on-device with all-MiniLM-L6-v2, selecting MPS, CUDA, or CPU automatically; no external API calls.
Inside the build
Next
mini-github
View →