MemPalace hits 49K stars: local-first AI memory with 96.6% LongMemEval recall, zero API calls
thisguyknowsai · x · 2026-10-29
The open-source MemPalace project has reached 49K GitHub stars with a local-first AI memory system that claims to outperform paid alternatives. Instead of the common "junk drawer" approach of dumping everything into semantic search, it organizes memory into a structured palace: people and projects become wings, topics become rooms, and original text lives in drawers that can be searched individually.
Key facts:
- 96.6% retrieval recall on LongMemEval with zero API calls, zero LLM, zero cloud; 98.4% on unseen questions with the hybrid pipeline, fully reproducible from the repo
- Strict verbatim storage — nothing is summarized or rewritten
- 29 MCP tools for palace reads, knowledge graph operations, and cross-wing search
- Pluggable retrieval layer (default ChromaDB); no data leaves your machine unless you opt in
More from coding & agent
- qwen-code PR Sizes All Qwen4 Models at 1M-Token Context Window — haider1 · 2026-10-29
- AI Engineer: Make Custom Evals a First-Class Citizen, Not Bigger Models — paulabartabajo_ · 2026-10-29
- Running 55 Claude Code subagents in one workflow for batch search tasks — AlchainHust · 2026-10-29
- Free DataTalks.Club x Nebius course: production AI agents for generative media — Al_Grigor · 2026-10-29
- MCP server makes Claude/ChatGPT cite datasheet pages for every part value — Winter-Drop-1334 · 2026-10-29
- Two devs used Claude to build a browser-based PS2 emulator, TabStationToo — yousriii · 2026-10-29