MemPalace is an AI memory system that stores conversation history as verbatim text and retrieves it using semantic search.
The tool solves the problem of maintaining accessible long-term context for AI agents without summarization or paraphrasing. It stores original content verbatim and organizes the index hierarchically—people and projects become wings, topics become rooms, and original content lives in drawers—allowing searches to be scoped rather than run against a flat corpus. The retrieval layer is pluggable, with ChromaDB as the current default backend; alternative backends can be integrated by implementing the interface defined in mempalace/backends/base.py. All data remains local by default unless explicitly opted into cloud storage.
Adoption suits projects where preserving exact conversation history matters and where semantic search over structured memory is valuable. The tool is designed for local-first operation with zero API calls required for core functionality. It performs at 96.6% R@5 raw on LongMemEval benchmarks. The project provides three skills for agent-guided setup and operation: mempalace for installation and operations, mempalace-recall for search-before-answer retrieval, and mempalace-task for logstream delegation. Installation is recommended via uv or pipx to isolate dependencies from system Python environments. Native Android support is not available, though Linux ARM64 users can run the tool in a Debian PRoot container on Termux.
The project maintains active engagement with users through documented security concerns, including a caution about impostor sites and a guide for Claude Code session retention. It provides multiple installation paths—agent-guided setup, direct CLI installation, and Docker—reflecting responsiveness to different user preferences. The codebase includes architecture documentation at mempalaceofficial.com/concepts/the-palace and maintains a detailed history of security incidents in docs/HISTORY.md.
