MemPalace
Conversation memory for agents, in Python, MIT licensed. It keeps the text of your sessions word for word and only chooses at search time, through semantic search: nothing is summarised or reworded on write. Storage follows the memory palace metaphor: one wing per person or project, one room per topic, and the original content in drawers, so you search within a scope instead of the whole corpus. For Claude Code it plugs in through an MCP server, hooks that save during the session and right before a compaction, and three setup and recall skills. Default storage is ChromaDB, swappable for SQLite, Qdrant, Milvus or Postgres. Alongside it, a temporal entity graph in SQLite records what is linked to what and since when, with relations that can be invalidated.
Strengths
- Nothing is lost on write: the original wording stays retrievable
- Search scoped by wing and room, so by project or topic
- Automatic save before compaction, where a session loses the most
Limitations
- A 45-tool MCP server carried into every session
- Everything is kept, including what was wrong one Tuesday: sorting is still on you
- Copycat sites around: install from GitHub or PyPI only
Best for
- Long work where finding the exact sentence from a past session matters more than its summary