Retrieval-Augmented Self-Recall — Part 6: The Fine-Tune That Did Nothing, and Shipping It as an MCP Server

Short summary

Part 6 of a series on Retrieval-Augmented Self-Recall. Fine-tuning all-MiniLM-L6-v2 on a 14-document corpus yielded exactly zero MRR/nDCG improvement because the base model already saturated it — but the same fine-tuning gave +0.24 MRR on a harder jargon corpus. The author ships the engine as recall_mcp, an MCP server so Claude can query its own memory with trust verdicts and calibrated confidence baked into structured output.

• •Fine-tuning embeddings gave zero improvement when base model already saturated the corpus, but +0.24 MRR on harder jargon corpus
• •Null result is valuable: it tells you when fine-tuning is unnecessary vs when it's needed
• •Engine ships as recall_mcp MCP server with honesty signals (trust verdicts, confidence, provenance) baked into structured output

Generated with AI, which can make mistakes.