ReMe is a memory management kit for AI agents that stores knowledge as a local-first, self-evolving personal knowledge base.
ReMe addresses the challenge of agents maintaining persistent, organized, and retrievable knowledge across conversations and sessions. It converts conversations and resources into structured Markdown files with frontmatter and wikilinks, treating memory as both readable files and queryable data. The approach combines BM25 search with optional embeddings and wikilink expansion to retrieve relevant passages at line-level granularity without requiring the entire knowledge base to be loaded into agent context. This design allows agents to progressively refine facts, preferences, procedures, and relationships over time while preserving source attribution.
The tool suits projects where multiple agents need to share a common knowledge workspace or where users want to maintain control over their agent's memory as ordinary files they can inspect, edit, and back up with standard tools. ReMe integrates with agent runtimes including QwenPaw and DeepSeek Harness through native integrations, SKILL.md, CLI, HTTP, MCP, or Python APIs. It distinguishes itself by storing memory durably as Markdown rather than in proprietary formats, allowing indexes and metadata to be rebuilt if needed, and by supporting both agent-driven knowledge evolution and direct user editing of the same workspace.
The project shows consistent development activity with regular commits and maintains documentation alongside the codebase. The tool requires Python 3.11 or later and is distributed through standard package management channels.
