agentscope-ai/ReMe

What it solves

ReMe provides a local-first, self-evolving personal knowledge base for AI agents. It solves the problem of agents lacking durable, long-term memory that is both human-readable and agent-accessible, allowing multiple agents to share a single workspace to retrieve, maintain, and evolve knowledge over time.

How it works

ReMe treats "memory as files," storing durable memory as ordinary Markdown files with frontmatter and wikilinks. It follows a capture-index-consolidate-recall loop:

• Capture: auto_memory distills conversation facts and auto_resource turns external files into daily cards.
• Index: auto_index creates searchable chunks, BM25 indexes, and wikilink graphs (with optional vector embeddings).
• Consolidate: auto_dream extracts reusable units from recent files to refine, create, or correct long-term "digest" nodes (personal facts, procedures, and wiki knowledge).
• Recall: Agents retrieve relevant line-level passages and their relationships using BM25, optional embeddings, and wikilink expansion.

Who it’s for

Developers building AI agents (such as those using DeepSeek Harness, OpenClaw, QwenPaw, or Claude Code) who need a persistent, local-first memory system that users can inspect and edit manually.

Highlights

• Human-Readable Memory: Memory is stored as Markdown files, making it easy for users to sync, back up, and edit with familiar tools.
• Self-Evolving: Progressively transforms conversations and resources into daily notes and long-term knowledge nodes.
• Cross-Agent Workspace: A single local workspace can be shared across different agent runtimes via CLI, HTTP API, MCP, or Python APIs.
• Hybrid Search: Combines BM25 and optional vector embeddings using reciprocal rank fusion (RRF) for precise, context-aware recall.
• Integrated Studio: Includes a web-based UI (ReMe Studio) for browsing, editing, and searching the workspace.