Post-Mortem: Building a Local MCP Server for Codebase Memory using Ollama and ChromaDB

Short summary

This post-mortem details building a local MCP server for codebase memory using Ollama and ChromaDB, motivated by cloud API billing unpredictability and privacy concerns. The author benchmarks mistral:7b (6.14s mean) against ornith:9b (13.39s mean) on an RTX 3050 8GB, finding mistral fits cleanly in VRAM while ornith spills into shared memory. Live workspace testing shows ornith produces better-structured synthesis with inline citations despite slower latency.

• •Local MCP server runs entirely on-device via Ollama—no cloud calls, no data leaves machine
• •mistral:7b averages 6.14s vs ornith:9b at 13.39s on RTX 3050 8GB VRAM
• •ornith:9b produces richer, better-cited synthesis but spills into shared memory on consumer GPUs

Generated with AI, which can make mistakes.