AI Native Landscape

mempalace

Tracked

Local-first AI memory that stores conversation history verbatim and retrieves it with semantic search — 96.6% Recall@5 raw on LongMemEval with zero API calls — over a pluggable retrieval backend.

AuthorMemPalaceOpen Sourced2026-04-05Last CommitUnknown

Overview

MemPalace is a local-first AI memory system that stores conversation history as verbatim text and retrieves it with semantic search — it does not summarize, extract, or paraphrase. The index is structured like a palace: people and projects become wings, topics become rooms, and original content lives in drawers, so searches can be scoped rather than run against a flat corpus. Nothing leaves your machine unless you opt in.

Key Features

  • Verbatim storage: conversation history is kept as original text, never summarized or paraphrased
  • Structured index with wings (people/projects), rooms (topics), and drawers (original content) for scoped search
  • Pluggable retrieval backend with ChromaDB as the default; alternatives drop in without touching the rest of the system
  • 96.6% Recall@5 raw on LongMemEval with zero API calls
  • Ships as a CLI, an MCP server, and three agent skills (mempalace, mempalace-recall, mempalace-task)
  • Runs via uv/pipx, in a venv, or as a multi-arch Docker image (amd64 + arm64)

Use Cases

  • Give Claude Code persistent memory across sessions that expire or get compacted
  • Recall past decisions and conversations with search-before-answer semantics before replying
  • Run a shared "palace" as a memory hub for multiple agents while keeping all data local

Technical Details

  • Storage is verbatim text; retrieval is semantic search over a structured index whose shapes (wings/rooms/drawers) let queries be scoped to a person, project, or topic instead of scanning a flat corpus
  • The retrieval layer is an interface defined in mempalace/backends/base.py; ChromaDB is the current default and any alternative backend can be substituted without changes elsewhere
  • The MCP server runs over stdio (docker run -i for the container flavor), and all state — palace, config, and the cached embedding model — persists under /data so a mounted volume survives restarts
  • Agent-guided setup via npx skills add MemPalace/mempalace detects the system, installs the Python package, configures MCP, and records whether the runtime came from uv tool, pipx, or pip; optional release checks contact only PyPI and never auto-install updates