ReMe: Build a File-Based Memory Workspace for AI Agents

Overview

ReMe is a Python memory-management kit for AI agents, built around a local-first, self-evolving personal knowledge base whose durable memory is ordinary Markdown. It addresses the gap between short-lived conversations and useful long-term context: agents can capture conversations and resources, progressively refine them into knowledge, and retrieve relevant passages without loading the whole workspace into context.

The same workspace can serve personal assistants and coding agents through the CLI, HTTP, MCP, Python APIs, and integrations for QwenPaw, DeepSeek Harness, OpenClaw, Claude Code, Hermes, and Codex. ReMe combines BM25 search with optional embeddings and wikilink expansion; embeddings are disabled by default.

Key Features

• Stores durable memory in editable Markdown files with frontmatter and wikilinks, while keeping indexes and metadata rebuildable.
• Captures conversation records and distills useful information into daily memory cards.
• Imports external resources into source-linked daily cards, including image resources.
• Consolidates daily notes into long-term knowledge and procedural memory through Auto Dream.
• Searches at passage level with BM25, wikilink expansion, and optional embedding-based retrieval.
• Offers workspace operations through a CLI, HTTP service, MCP server, Studio interface, and Python API.
• Supports integrations that add recall and memory capture to agent workflows.

Use Cases

• Personal assistant users can accumulate preferences, facts, and recurring context across conversations while retaining direct control of the files.
• Coding-agent users can preserve project procedures and prior decisions for retrieval in later sessions.
• Researchers can turn reference materials into linked, searchable notes alongside conversation-derived knowledge.
• Agent developers can add shared memory to supported runtimes through ReMe’s service interfaces or Python API.

Getting Started

Install the core package with Python 3.11 or newer, then start the service:

pip install "reme-ai[core]"
reme start

See the README for workspace setup, configuration, Docker deployment, and agent integrations.

Alternatives

• EverOS: EverOS focuses on persistent conversation memory in editable Markdown, while ReMe also captures resources and refines durable knowledge.
• memoripy: Memoripy tracks changing facts with evidence screening and cited recall, rather than centering a Markdown knowledge base.
• OpenViking: OpenViking organizes agent context in a browsable virtual filesystem and supports hosted use, while ReMe centers on local Markdown files.
• cabinet: Cabinet adds scheduled jobs and team workflows to its Markdown knowledge base, beyond ReMe's agent memory-management focus.

Considerations

• The project is active, with regular commits and releases, and its recent issue and pull-request handling has been brisk. Its contributor base is distributed rather than dominated by one person, but the feature set and integrations are substantial, so evaluate the specific workflows your deployment needs.
• Python 3.11 or newer is required. Docker and Compose are also supported; the README does not specify a GPU requirement.
• Basic file operations and BM25 search can work without model credentials. LLM credentials are needed for memory evolution workflows, while embedding retrieval requires separately enabling embedding components and configuring an embedding API.
• ReMe’s durable workspace files are user-managed, but the service and integrations still need to be configured for your environment. Review the deployment and security guidance before exposing the service beyond its intended boundary.
• The Apache-2.0 license permits broad use, subject to its license terms and notices.