DOCSEARCHbyAlgolia
Submit Docs

OpenClaw Orchestration Framework

website·Indexed 30 days ago

https://eaveluo.com

An open-source multi-agent orchestration framework featuring agent isolation models, message routing via a Bindings system, and a Sub-Agents system with hierarchical session keys. It implements the Agent Collaboration Protocol (ACP) for editor integration and utilizes the QMD memory backend with a three-stage hybrid search process (BM25, vector search, and LLM reranking) to prevent agent forgetfulness. Version 4.5 introduces the Dreaming feature for machine memory processing.

Tokens
22.9K
Snippets
116
Records
168
Agent score
49%

What's inside OpenClaw

  1. Overview of QMD Memory Backend for OpenClaw

    QMD (Query Markup Documents) serves as a local memory exoskeleton for AI Agents, solving retrieval failure issues that occur as logs accumulate. It employs a three-stage hybrid search pipeline to ensure high precision and semantic understanding:

    1. BM25 Full-Text Retrieval: Fast keyword matching, ideal for code, IDs, and specific terms.
    2. Vector Semantic Search: Uses Jina v3 embedding models (GGUF format) to find conceptually similar content regardless of exact wording.
    3. LLM Reranking: Uses a local Qwen3 0.6B model to evaluate the top candidates from the first two stages (merged via Reciprocal Rank Fusion) and select the most relevant final results.

    This architecture runs entirely locally, ensuring zero API costs and data privacy.

  2. Overview of OpenClaw 4.5 Dreaming Mechanism

    The Dreaming mechanism in OpenClaw 4.5 is an automated memory management system that converts fragmented short-term interactions into long-term memory. It replaces manual maintenance of memory files (like MEMORY.md) by implementing a three-stage process inspired by human sleep: Light, REM, and Deep. This allows the AI to categorize, evaluate, and consolidate memories without user intervention.

  3. Overview of QMD Hybrid Search for OpenClaw

    QMD (Query Markup Documents) serves as a local memory exoskeleton for AI agents, solving retrieval failures that occur as logs accumulate. It uses a three-stage hybrid search pipeline to ensure high-precision retrieval:

    1. BM25 (Keyword Matching): Fast, precise matching for specific keywords, code, or IDs. It excels at finding documents containing exact search terms.
    2. Vector Search (Semantic Meaning): Uses the Jina v3 embedding model (GGUF format, ~300MB) to find conceptually similar content, even if the wording differs.
    3. LLM Reranking (Quality Gate): Uses a local Qwen3 0.6B model to score the top candidates from the first two stages (merged via Reciprocal Rank Fusion - RRF), returning the top 6 most relevant snippets (up to 700 characters each).

  4. Overview of OpenClaw 4.5 Dreaming mechanism

    The Dreaming mechanism in OpenClaw 4.5 is an automated memory management system that simulates human sleep to transition fragmented short-term interactions into long-term memory. It replaces manual maintenance of memory files by automatically revisiting, categorizing, and evaluating daily conversations to decide what should be persisted in the official archive (MEMORY.md).

  5. Overview of the Agent Collaboration Protocol (ACP)

    The Agent Collaboration Protocol (ACP) is an open standard initiated by Zed Industries to enable seamless integration between AI agents and editor environments. It solves the 'three-app problem' (switching between editor, AI agent, and browser) by decoupling the agent from the editor. This allows any ACP-compatible editor to work with any ACP-compatible agent, ensuring no vendor lock-in and allowing both components to evolve independently.

  6. Overview of the Agent Collaboration Protocol (ACP) Specification

    The Agent Collaboration Protocol (ACP) is a JSON-RPC 2.0 based communication protocol designed to decouple AI agents from editor environments using a bridge pattern. It provides a standardized way to manage agent lifecycles and tool integration, supporting core operations such as spawning, steering, and killing agents. The protocol is designed to integrate with mainstream tools like Claude Code and Codex.

  7. Overview of the Agent Collaboration Protocol (ACP)

    The Agent Collaboration Protocol (ACP) is an open standard initiated by Zed Industries to enable seamless integration between AI agents and editor environments. It solves the 'three-app problem' (constant switching between editor, AI agent, and browser) by decoupling the agent from the editor. This allows any ACP-compatible editor to work with any ACP-compatible agent, preventing vendor lock-in and allowing both components to evolve independently.

  8. ACP Protocol Architecture and Transport

    ACP is built on JSON-RPC 2.0 and utilizes request-response and notification communication patterns. The protocol stack consists of three layers:

    1. Application Layer: Handles session management (initialize, prompt, update, etc.).
    2. JSON-RPC 2.0 Layer: Manages requests, responses, notifications, error handling, and batch requests.
    3. Transport Layer: Primarily uses stdio, but also supports WebSockets, Unix sockets, and TCP sockets.

  9. Understand OpenClaw Agent Isolation Model

    OpenClaw uses a fully isolated agent model where each agent acts as an independent 'brain' with its own dedicated resources to prevent credential conflicts and state leakage. Isolation is applied across four key areas:

    • Workspace Isolation: Each agent has a unique directory containing identity files like AGENTS.md, SOUL.md, and USER.md.
    • Authentication Isolation: auth-profiles.json is separated per agent.
    • Session Isolation: Sessions are stored independently under ~/.openclaw/agents/<agentId>/sessions.
    • Skill Isolation: Agents load exclusive skills from their own skills/ directory, though they can access shared skills in ~/.openclaw/skills.

  10. Understand the Dreaming memory architecture for AI agents

    Dreaming is a background memory consolidation process designed to transform AI assistants from short-term chat tools into long-term agents. Unlike simple retrieval-augmented generation (RAG) or long-context windows—which can be 'noisy' and include outdated information—Dreaming acts as an anti-entropy mechanism. It periodically synthesizes, compresses, updates, and forgets information to ensure the agent maintains a current and executable context of user preferences, project details, and historical facts without requiring the user to explicitly manage a database.

  11. Configure OpenClaw QMD memory backend

    To enable high-recall semantic search in OpenClaw, configure the memory backend to use qmd. This setup allows the agent to handle complex queries and fuzzy concept matching better than standard SQLite.
    {
  "backend": "qmd",
  "citations": "auto",
  "interval": "5m",
  "waitForBootSync": false,
  "maxResults": 6,
  "maxChars": 700
}

  12. ACP Protocol Architecture and Transport

    ACP is built on the JSON-RPC 2.0 specification, utilizing request-response and notification communication modes. The protocol stack consists of three layers:

    1. Application Layer: Handles session management (initialization, prompts, updates, etc.).
    2. JSON-RPC 2.0 Layer: Manages request/response, notifications, error handling, and batch requests.
    3. Transport Layer: Supports multiple transport mechanisms, with stdio being the primary method, alongside WebSocket, Unix socket, and TCP socket.