Scoring memories without embeddings

● finished

Setup

Thirty saved memories, twelve retrieval tasks, and a hand-written relevance key.

What I tried

Each memory received scores for recency, declared topic, prior reuse, and confidence. I compared the combined rank with a semantic-search baseline.

Result

The metadata baseline won on recurring project decisions and lost on loosely phrased conceptual questions.

Lesson

The two approaches fail differently. A hybrid is more promising than tuning either in isolation.

Next step

Add a lightweight semantic score and test on one hundred less curated memories.