MemPalace

MemPalace stores and retrieves memory for AI workflows on your own machine and organises it spatially — wings for people, projects or topics, rooms for days, sessions or threads — an explicit nod to the method of loci. It publishes a recall figure on LongMemEval, requires no API calls, and ships MCP tooling so an assistant can query it.

Two things are worth separating here. The local-first stance is a genuine property: memory about your work is among the most revealing data an assistant touches, and keeping retrieval on the machine removes a whole category of question. The organising metaphor is the part to examine sceptically — a spatial structure is a retrieval scheme, and its value depends on whether the thing you are looking for is where you would have filed it, which is exactly the judgement people are bad at about themselves. A published benchmark number is a starting position, not a result: LongMemEval measures recall over a defined task, and the honest question is what the failure cases look like when the memory is yours.