Show HN: ClawMem – Open-source agent memory with SOTA local GPU retrieval

Show HN: ClawMem — open-source agent memory with SOTA local GPU retrieval. Shares a single SQLite vault between Claude Code and OpenClaw (no syncing). Runs on Bun, WAL SQLite (FTS5+vec0). Repo: https://github.com/yoloshii/ClawMem

Retrieval stack is assembled from recent papers/projects: QMD pipeline (BM25+vector+RRF+query expansion) + cross-encoder reranker, plus SAME/MAGMA/A‑MEM ideas. Run via llama-server or node-llama-cpp — or zembed-1 + zerank-2 for higher quality (~12GB VRAM).

Tradeoffs: author reports noticeably better retrieval with zembed+zerank but it needs ~12GB VRAM; fallback is cloud embeddings. Codebase ≈19K LOC; main store is a 4K-line god object. Reality check: vault content dictates usefulness — three files = thin memory.

Takeaway for builders/agencies: if you want local, persistent agent memory that both CLI and voice agents can use, ClawMem is the most complete open option I’ve seen. Quick test: index one project’s docs vs full vault and compare context hits. Anyone tried zembed+zerank locally