MemPalace Organizes Chats Into Local Memory Graph

The core idea is interesting and practical, but the hype is stronger than the evidence.

• –The main differentiator is structural organization, not a magic new retrieval model: MemPalace stores verbatim conversations locally and layers semantic search plus palace-style grouping on top.
• –The strongest claim is the raw-mode LongMemEval result, which the repo says is 96.6% R@5; the experimental AAAK compression mode scores lower, so the headline is really about fidelity-first storage.
• –If you already use local-file RAG or a conventional memory layer, the likely win here is better long-horizon recall and easier navigation across related conversations, not a fundamentally new memory primitive.
• –The project being free, open source, local, and MCP-friendly makes it easy to try, but I would still want independent benchmark runs and real workflow comparisons before buying the “best ever” framing.