agentscope-ai/ReMe

ReMe: Memory Management Kit for Agents - Remember Me, Refine Me.

What it solves

ReMe provides a local-first, long-term memory layer for AI agents, allowing them to turn conversations and external resources into a structured, searchable, and editable knowledge base. It solves the problem of agents forgetting information across sessions by creating a persistent memory system that both humans and agents can read and write directly.

How it works

ReMe treats "memory as files," using Markdown files with frontmatter and wikilinks to store information. It follows a capture-index-consolidate-recall loop:

  • Capture: Raw conversations (session/) and external resources (resource/) are processed into daily memory cards (daily/).
  • Index: A background process maintains BM25, wikilink graphs, and optional embedding indexes for retrieval.
  • Consolidate: An "auto-dream" process distills stable knowledge from daily cards into long-term memory nodes (digest/) categorized by personal facts, procedures, and general wiki knowledge.
  • Recall: Agents retrieve information using a hybrid search (BM25, embeddings, and wikilink traversal).

Who it’s for

  • AI Agent Developers: Those building personal assistants (e.g., QwenPaw, Hermes) or coding agents (e.g., Claude Code) who need a persistent, human-readable memory layer.
  • LLM Application Builders: Users wanting to turn their AI interactions and notes into a linked Markdown knowledge base.
  • Self-evolving Agents: Developers creating agents that learn from experience by saving successful paths and reflections.

Highlights

  • Human-Readable Memory: Memory is stored as ordinary Markdown files, making it directly editable by users.
  • Self-Evolving Knowledge Base: Automatically transforms raw data into long-term memory via auto_memory, auto_resource, and auto_dream jobs.
  • Hybrid Search: Combines keyword matching (BM25), semantic recall (embeddings), and relationship expansion (wikilinks).
  • Agent-Friendly Integration: Supports multiple integration paths including a Python SDK, HTTP API, MCP server, and CLI skills.