MemPalace

MemPalace is an open-source, local-first memory layer for LLM agents that stores conversation history verbatim and retrieves it through semantic search. Instead of summarizing or paraphrasing (a common failure mode for chat-memory systems that quietly lose fidelity over time), it keeps original content intact and organizes it into a metaphorical palace: people and projects become 'wings', topics become 'rooms', and the raw text lives in 'drawers'. That structure lets retrieval be scoped to a wing or room rather than run against a flat corpus, which is how it reports 96.6% R@5 raw on LongMemEval with zero API calls. Everything runs on your own machine with ChromaDB as the default vector store, and pluggable backends cover SQLite, Milvus, Qdrant, and pgvector for teams that already have infrastructure. On top of storage it ships an MCP server exposing 36 tools for palace operations, temporal knowledge-graph management (entity-relationship edges with validity windows), and per-agent diaries, plus auto-save hooks for Claude Code, Codex CLI, and Cursor so context accumulates in the background as you work. Specialised subagents get their own isolated wing and diary, which keeps a coding assistant's memory from bleeding into a research agent's. Embeddings use either embedding-gemma-300m for multilingual work or all-MiniLM-L6-v2 for a lightweight English-only footprint, and no LLM is required for the retrieval path itself. Install is via `uv tool install mempalace` or `pipx install mempalace`, with Docker images for CPU and GPU. Target audience is engineers building persistent-memory agents, MCP tool developers, and privacy-conscious teams who want conversation memory without cloud dependencies or per-token bills.