HyperspaceDB v3.1.4: 1-bit ADC gets 107× search speedup, ships Mem0 drop-in replacement

Sam_YARINK · reddit · 2026-09-03

HyperspaceDB v3.1.4 targets agent memory and vector search: Turbo 4-bit Lloyd-Max quantization delivers 7.8× RAM compression at 96.4% Recall@10, while 1-bit rotated ADC cascades hit 107× raw search speedup (99.9% Recall@10 with a two-pass rerank, 15–20× net throughput). A hyperspace-memory package is a 100% API-compatible Mem0/Zep drop-in (Python/TS) claiming <0.5ms latency and 98% storage savings. It also adds multi-step agent trajectory tracking with a Lyapunov stability exponent to detect hallucinations and reasoning loops, plus a zero-config MCP memory server (npx -y mcp-hyperspace-memory) for Cursor, Claude Desktop, and Windsurf.

Original post →

More from coding & agent

coding & agent channel →