MindSlice

Portable memory for AI agents, with isolation and evolution built in.

Keep project knowledge, personal preferences and shared handbooks in plain Markdown. Mount what an agent needs, let it learn in its own workspace, and bake those learnings into reusable versions. Connect through MCP, the CLI or Python.

Why MindSlice?

Isolation. A slice is a folder of Markdown cards you can mount, unmount and share independently. Mounted slices stay read-only through the engine; new learning goes into a separate overlay. Multiple workspaces can reuse the same baseline while keeping their own changes. Pending edits stay outside live memory until approved.

Evolving memory. Cards can explicitly extend or supersede earlier knowledge, so retrieval follows recorded relationships under your trust rules. bake turns selected changes into new slice versions, preserving the original folders and returning the new paths.

Hot/cold memory focuses on what stays in context and what is retrieved later. LLM Wiki focuses on compiling sources into a maintained knowledge base. MindSlice makes memory boundaries and controlled evolution first-class engine operations. How these approaches fit together →

Python 3.12+. No runtime dependencies, database server or LLM API calls inside the engine.

Get started

Paste this into your coding agent:

Clone https://github.com/pancerZH/MindSlice.git (or use the existing checkout).
Read skills/mindslice-setup/SKILL.md to finish setup.

The agent helps you choose a workspace, connect your app, and choose how to start:

Forge is an optional document-conversion tool. It creates a new workspace using an agent harness; MindSlice handles storage, retrieval, mounting and memory evolution without needing Forge or an LLM. The builder skill guides conversion and integration with an existing workspace.

Prefer manual setup? With Python 3.12+ and pipx installed:

git clone https://github.com/pancerZH/MindSlice.git
cd MindSlice
pipx install -e .
mindslice init ~/mindslice

Then connect your agent and tell it what to remember. Run mindslice -w ~/mindslice doctor, then verify a new session. Full setup guide · Review changes and bake slices

Documentation

• Concepts — isolation, evolution, and how MindSlice fits other memory architectures.
• Reference — commands, Python API, MCP tools and card formats.
• History and recovery — explain memory, roll back slices, back up and restore workspaces.
• Build slices with Forge — convert documents and connect the resulting workspace or slices.
• Maintain memory — consolidate knowledge and repair relationships.
• Evaluations · Design · Uninstall

License

Apache License 2.0. Benchmark excerpts quoted in the evaluation records keep their datasets' licences; see NOTICE.