# zerikai_memory

zerikai_memory is a standalone, local-only Python MCP server that gives any IDE persistent, workspace-isolated memory of a codebase's entities, built with deterministic tree-sitter parsing rather than an LLM-driven index.

The server parses a codebase deterministically with tree-sitter to capture entities such as functions, classes and docstrings, and stores them with deep code descriptions in a local ChromaDB vector store, scoped to one workspace at a time. Retrieval combines vector similarity with lexical re-ranking and strict source verification, citing the entity, file and line number behind each result, aimed at reducing the token cost of feeding an agent raw file contents.

An optional layer called Jev, from TypeSafe AI, adds a second, narrower judgment model that scores retrieved passages for relevance, evidence and contradiction and flags prompt-injection attempts, producing a short evidence report the agent can act on; it is off by default and the rest of the pipeline runs unchanged if it is unavailable.

It is written in Python and pairs by design with low-cost DeepSeek API usage, and is released under the MIT license.

Website: https://github.com/KikeVen/zerikai_memory