ReMe: Memory Management Kit for Agents
Turns conversations and resources into readable, editable, searchable Markdown memory for AI agents.
Evidence: README explicitly states embeddings are disabled by default and basic operations don't require LLM credentials, reflecting least privilege; but no permission model or security audit provided. User confirmation: CLI commands like write/edit execute directly without confirmation prompts, but docs suggest user-editable files, partially reflecting user control. Data flow transparency: README details directory structure and data flow, but doesn't explain how data is accessed by external services. Sensitive data handling: docs mention session and resource files, but no encryption or anonymization measures. Dependency security: pyproject.toml lists dependencies, but no vulnerability scanning or lock files. External effects: service binds to 127.0.0.1 by default, but network exposure risks not discussed. Rollback: no backup or rollback mechanism mentioned. Source attribution: README and LICENSE clearly attribute copyright, but publisher unverified. Deductions: lack of security details and user confirmation mechanisms.
Evidence: README and test files are consistent, showing good self-consistency; but no failure recovery or degradation strategies. Dependency availability: dependency list complete, but no version compatibility or mirrors. Failure messages: test files include timeout error messages, but user docs lack common error handling guide. Deductions: missing failure handling and error documentation.
Evidence: README clearly identifies target users (personal assistants, coding agents, etc.) and use cases, and provides multiple integration paths (CLI, MCP, SDK), showing strong adaptability. Capability boundaries: docs distinguish features requiring LLM credentials from those not, boundaries clear. Trigger precision: commands like auto_memory have clear trigger conditions, but no details on frequency or conditions. Environment fit: supports Python 3.11+, cross-platform, but no specific OS limitations. Deductions: insufficient trigger details and environment constraints.
Evidence: README structure is clear, with installation, usage, integration sections, good information architecture. Install notes detailed, including pip and source. Naming stability: version history mentioned, but no API stability guarantee. Examples and FAQ: 5-minute demo and cookbooks provided, but no FAQ. Known limitations: not explicitly listed. License: Apache-2.0 complete. Versioning and changelog: no CHANGELOG provided. Maintenance responsibility: README provides contribution guide and community support, but no explicit maintainers. Deductions: missing FAQ, known limitations, and changelog.
Evidence: Output is Markdown files, directly editable and searchable, good usability. Marginal value: as a memory layer, provides unique features, but needs comparison with existing tools. Cost-benefit: requires LLM credentials for core features, but basic features free, no performance benchmarks. Deductions: missing performance data and cost analysis.
Evidence: Claims in README (e.g., features) are supported by docs and tests, traceable. Cross-source corroboration: test files provide integration tests, but no comparison with other sources. Fact-inference separation: docs distinguish feature descriptions and design philosophy, but no explicit inference labeling. Deductions: lack of external validation and explicit fact/inference marking.
- Publisher identity unverified; be cautious about supply chain risks.
- Dependencies not pinned, potential security vulnerabilities.
- Sensitive data handling lacks encryption or anonymization measures.
- No rollback mechanism; risk of data loss.
- LLM credentials used for core features; ensure secure storage.
What does this agent do, and when should you use it?
ReMe is a local-first memory layer for AI agents that transforms conversations and resources into file-based long-term memory. It uses Markdown files with frontmatter and wikilinks as memory nodes that both users and agents can read and write directly. Core features include auto memory, auto resource, auto dream, and proactive topics, which progressively convert raw data into long-term memories while building wikilink relationships. It offers a hybrid retrieval approach combining wikilinks, BM25, and optional embeddings. ReMe provides multiple integration paths: CLI, HTTP API, MCP server, and Python SDK, making it compatible with agents like QwenPaw and Claude Code. It also includes optional cookbooks for daily paper analysis and financial research. The project is Apache-2.0 licensed and requires Python 3.11+.
ReMe runs as a local service, exposing a CLI with jobs such as start (launch service), write/read/edit (manage Markdown memory files), search (hybrid retrieval with BM25 and wikilinks, plus optional vectors), auto_memory (extract daily memory cards from conversations), auto_resource (convert resource files into daily cards), auto_dream (consolidate daily cards into long-term memory under digest/), proactive (generate topics for the agent to mention), and reindex (rebuild indexes). It also provides an MCP server for Claude Code integration and a SKILL.md for CLI-capable agents. The workflow captures raw sessions and resources, processes them into daily/ cards, and distills them into digest/ long-term memory nodes. The system maintains indexes, wikilink graphs, and optional embedding stores in metadata/.
- Personal assistants like QwenPaw, OpenClaw, or Hermes need a user-editable long-term memory layer to remember preferences and past interactions.
- Coding agents like Claude Code need to preserve coding style, project context, repository decisions, and workflow experience across sessions.
- Researchers want to build a searchable, traceable, linked Markdown knowledge base from papers, notes, and resources.
- Developers want to equip agents with the ability to learn from experience by saving successful paths, failed attempts, and reusable procedures.
What are this agent's strengths and limitations?
- Local-first and file-based: memory is stored as Markdown, making it transparent and directly editable by users and agents.
- Progressive hybrid search: combines wikilinks, BM25, and embeddings for effective retrieval across keywords, semantics, and relationships.
- Multiple integration paths: CLI, HTTP API, MCP server, and Python SDK support various agent runtimes.
- Self-evolving knowledge base: auto_memory, auto_resource, and auto_dream automatically transform conversations and resources into long-term memory with wikilink relations.
- Requires Python 3.11+, which may limit some deployment environments.
- Automatic memory features depend on external LLM APIs, requiring API keys, network access, and incurring potential costs.
- Embedding-based retrieval is disabled by default and requires manual configuration changes to enable.
- No detailed performance or scaling guidance is provided, which may be a concern for large-scale memory workloads.
How do you install or deploy this agent?
Install with Python 3.11+: pip install "reme-ai[core]". Or from source: git clone https://github.com/agentscope-ai/ReMe.git && cd ReMe && pip install -e ".[core]". Set environment variables: LLM_API_KEY and LLM_BASE_URL (required for auto_memory, auto_resource, and auto_dream); optionally EMBEDDING_API_KEY and EMBEDDING_BASE_URL for embedding-based retrieval.
How do you use this agent?
Start the service: reme start (default port 2333). Write a memory node: reme write path=digest/wiki/quick-start-demo name="Quick Start Demo" description="A first ReMe memory node" content="...". Search: reme search query="agent memory markdown" limit=5. Read: reme read path=digest/wiki/quick-start-demo start_line=1 end_line=20. To enable embeddings, uncomment as_embedding and embedding_store in reme/config/default.yaml and set EMBEDDING_API_KEY.
How does this agent compare with similar options?
The repository mentions MemoryScope as a previous version and recommends integrations with QwenPaw, OpenClaw, and Hermes. No direct comparison to other memory tools is given.