Three Show HN Posts Ditch Vector Databases for Agent Memory

Three agent-memory projects landed on Show HN within the same six-hour window on August 7. Each one cuts the vector database out of the stack.

None of the three has real traction, sitting at one to four points on Show HN. The stakes fall on anyone maintaining a vector-backed memory stack. If a flat SQLite file does the job, that infrastructure is weight the project didn't need.

Remembrane bills itself as agent memory in one SQLite file, zero dependencies, per its GitHub repo. llmem offers local persistent memory for AI coding with no embeddings. XSAF, an Extra Small Agent Framework, carries the same instinct in its name.

The target is implicit but obvious from what got cut: no vector store, no dependencies, no framework, and in llmem's case, no embeddings at all. The stack that made memory-heavy agent projects the ones topping star charts earlier this year is exactly what these three strip out.

What to watch

The bet is that most agents don't need a vector database, just a file they can read and write. If Remembrane, llmem, and XSAF stay stuck at a handful of points while the heavyweight memory stacks keep adding stars, the subtraction bet loses. If builders start picking them over the bigger stacks instead, the complexity in those older projects was never load-bearing.