DiffMem drops heavy retrieval stack

This is the kind of rewrite that sounds almost silly until you realize the problem was never “more retrieval,” it was “too much middleware around primitives the model already understands.”

• –Git history is genuinely useful for temporal questions that embeddings miss, especially co-occurrence across sessions and how relationships evolve over time.
• –The old stack carried a real ops cost: sentence-transformers, BM25, sklearn, and numpy bloated the container and made Cloud Run less reliable.
• –The new design is cleaner, but it shifts risk into repo hygiene, shell safety, and prompt discipline, so it works best when the memory repo is well structured.
• –“Return pointers, not content” is the strongest idea here, because it keeps the model’s context lean and pushes expensive fetching into code.
• –If this holds up broadly, it’s a nice argument that a lot of agent infrastructure should borrow more from version control and less from bespoke retrieval stacks.