Open-Claw on emsenn.net

OpenClaw Continuous Learning and Behavioral Evolution

Sat, 04 Apr 2026 00:00:00 +0000

¶Overview

OpenClaw achieves continuous learning and behavioral evolution through a three-layer architecture that persists knowledge across sessions, enables agents to autonomously extend capabilities, and filters learning opportunities through managed input gating. Unlike stateless chatbots that reset after each conversation, OpenClaw agents accumulate durable context, self-modify through skill creation, and evolve their decision boundaries through policy mechanisms.

¶Memory Persistence Layer

OpenClaw treats memory as the primary mechanism for accumulation across sessions. The memory system is Markdown-based and file-first, making it auditable and portable.

ClawHub

Mon, 09 Mar 2026 00:00:00 +0000

ClawHub is the public skill registry for OpenClaw. It provides a discovery and distribution surface for versioned skill bundles.

¶Role

The official docs position ClawHub as the normal way to search, install, update, and publish OpenClaw skills. Its CLI supports search, installation, updates, and sync-based publishing.

¶Function

ClawHub turns skills into portable, shareable artifacts rather than only local folders. It therefore extends OpenClaw’s skill system with a public registry layer.

Companion Tools for OpenClaw, March 2026

Mon, 09 Mar 2026 00:00:00 +0000

¶Purpose

This text surveys the tools that most often recur around OpenClaw in current official documentation and practical guide material as of March 9, 2026. The source base does not support exact usage counts, so “often used” here means recurring companions in the current OpenClaw ecosystem rather than measured market share.

¶What the tool stack reveals

The current OpenClaw ecosystem is converging on a distinctive practice. OpenClaw itself stays as the long-running control plane, memory layer, and multi-channel agent surface. Specialized work is pushed outward:

OpenClaw ACP Agent

Mon, 09 Mar 2026 00:00:00 +0000

An OpenClaw ACP agent is an external harness runtime that OpenClaw can run through an ACP backend plugin. ACP agents let OpenClaw delegate work to tools such as Codex, Claude Code, OpenCode, Gemini CLI, or Pi without reducing them to native OpenClaw sub-agents.

¶Role

The official ACP docs distinguish ACP sessions from OpenClaw-native sub-agent runs. ACP is used when the operator wants an external runtime, its own session semantics, or harness-specific capabilities.

OpenClaw DM Policy

Mon, 09 Mar 2026 00:00:00 +0000

The OpenClaw DM policy is the configuration that determines how OpenClaw handles inbound direct messages before they reach an agent. It is a channel-level gate on message processing, not only a cosmetic notification preference.

¶Modes

OpenClaw documents four DM policy modes:

pairing: unknown senders receive a short code and their message is ignored until approval
allowlist: only approved senders can reach the bot
open: any sender can reach the bot if the channel allowlist explicitly opts in
disabled: inbound DMs are ignored

¶Function

This policy is one of OpenClaw’s primary safety controls for consumer messaging surfaces. It constrains who can steer a tool-using assistant before prompts, tool calls, or session routing become relevant.

OpenClaw Gateway

Mon, 09 Mar 2026 00:00:00 +0000

The OpenClaw Gateway is the long-running control-plane process in an OpenClaw deployment. It coordinates sessions, channels, tools, configuration, routing, and web interfaces while the assistant remains the user-facing product surface.

¶Role

The Gateway gives OpenClaw a persistent runtime. It keeps the assistant reachable through chat channels, the Control UI, WebChat, and paired device surfaces without reducing the system to a single interactive terminal session.

¶Trust boundary

OpenClaw’s security model treats the Gateway as a trusted operator boundary, not as a hostile multi-tenant service. Within one Gateway instance, authenticated control-plane access carries authority over the configured agents, sessions, and policies in that deployment.

OpenClaw HEARTBEAT.md

Mon, 09 Mar 2026 00:00:00 +0000

OpenClaw HEARTBEAT.md is an optional workspace file that defines what the assistant should check or notice when it is idle. It is the main configuration surface for low-frequency proactive behavior.

¶Role

The official heartbeat documentation describes HEARTBEAT.md as a small checklist that the default prompt reads on a regular heartbeat. Independent guides distinguish it from cron: cron handles scheduled tasks at exact times, while HEARTBEAT.md handles opportunistic awareness and gentle proactive suggestions.

OpenClaw Hook

Mon, 09 Mar 2026 00:00:00 +0000

An OpenClaw hook is a directory-discovered event handler that runs when command or lifecycle events fire inside the Gateway. Hooks are the event-driven automation layer of OpenClaw.

¶Structure

The official hooks docs describe a hook as a directory containing HOOK.md plus a TypeScript handler. Hooks are discovered from workspace, managed, and bundled directories, with workspace hooks having highest precedence.

¶Role

Hooks let operators attach automation to events such as /new, bootstrap, or gateway startup. The bundled session-memory, bootstrap-extra-files, command-logger, and boot-md hooks show the main current uses: memory capture, workspace augmentation, audit logging, and startup actions.

OpenClaw Memory

Mon, 09 Mar 2026 00:00:00 +0000

OpenClaw memory is the Markdown-based memory layer in an OpenClaw workspace. Its source of truth is written files on disk rather than session-local model state.

¶Structure

The official memory docs describe two default durable layers:

MEMORY.md for curated long-term memory
memory/YYYY-MM-DD.md for dated logs and running context

Memory search tools are provided by the active memory plugin, but the files remain primary. Search augments retrieval; it does not replace the underlying Markdown record.

OpenClaw Node

Mon, 09 Mar 2026 00:00:00 +0000

An OpenClaw node is a paired device surface that exposes host-local capabilities to an OpenClaw Gateway. Nodes let the assistant act through device-specific interfaces such as voice, camera, screen capture, notifications, and host-local command execution.

¶Pairing model

OpenClaw documents node pairing as an operator-approved connection to a Gateway. Once paired, node actions are treated as trusted operator actions on that device rather than as requests from an untrusted external sender.

OpenClaw Plugin Slot

Mon, 09 Mar 2026 00:00:00 +0000

An OpenClaw plugin slot is an exclusive plugin category where only one plugin can own a given runtime role at a time. Slots let OpenClaw keep its core lean while making selected extension points explicit and mutually exclusive.

¶Examples

OpenClaw documents memory and contextEngine as plugin slots. If multiple plugins declare the same exclusive kind, only the selected plugin loads for that slot and the others remain disabled.

OpenClaw Research and Math Assistant Patterns, March 2026

Mon, 09 Mar 2026 00:00:00 +0000

¶Purpose

This text surveys how current expert material suggests using OpenClaw as a research assistant and then extends that evidence into a cautious math-assistant pattern. The research-assistant material is directly grounded in sources. The math-assistant section is an inference from those sources because the surveyed expert writing contains little OpenClaw-specific mathematics guidance.

¶The research-assistant pattern

The clearest expert playbook comes from Carl Vellotti’s research workflow guide. Its structure is simple:

OpenClaw Skill

Mon, 09 Mar 2026 00:00:00 +0000

An OpenClaw skill is a directory-based instruction and tool bundle that teaches an OpenClaw agent how to perform a specialized task. The official docs define a skill as a directory containing a SKILL.md file and, optionally, scripts or other resources.

¶Loading model

The official skills documentation describes three main loading locations:

bundled skills shipped with OpenClaw
managed or local skills under ~/.openclaw/skills
workspace skills under <workspace>/skills

When names conflict, workspace skills take precedence over shared and bundled skills.

OpenClaw SOUL.md

Mon, 09 Mar 2026 00:00:00 +0000

OpenClaw SOUL.md is a workspace file that defines the assistant’s identity, style, values, and behavioral boundaries. In practice it is the main place where an operator specifies how the assistant should speak, what it should prioritize, and where it should stop.

¶Role

The official bootstrapping flow writes SOUL.md during first-run setup after collecting identity and preference information. Independent setup guides treat it as the highest-leverage file to write early because it shapes every later interaction.

OpenClaw Summary, March 2026

Mon, 09 Mar 2026 00:00:00 +0000

¶Purpose

This text surveys OpenClaw as of March 9, 2026. OpenClaw is a self-hosted personal AI assistant platform organized around a persistent OpenClaw Gateway that exposes one assistant across the channels and devices a user already uses.

¶What OpenClaw is

The OpenClaw repository describes the project as a personal AI assistant that runs on the user’s own devices. The README’s central claim is that the Gateway is only the control plane. The product is the assistant itself: an always-available agent that can answer through messaging platforms, web chat, and companion apps.

OpenProse

Mon, 09 Mar 2026 00:00:00 +0000

OpenProse is a markdown-first workflow format and OpenClaw plugin for explicit, reusable agent programs. In OpenClaw it ships as a plugin that installs a skill pack plus a /prose slash command.

¶Role

Programs live in .prose files and can spawn multiple sub-agents with explicit control flow. The official docs describe research and synthesis, incident triage, code review, and content pipelines as core use cases.

¶Scope

OpenProse sits above individual tools and below general agent behavior. It is the program layer for repeatable multi-step workflows.

QMD

Mon, 09 Mar 2026 00:00:00 +0000

QMD is a local-first search sidecar that OpenClaw can use as an experimental memory backend. In this mode, Markdown remains the source of truth while OpenClaw shells out to QMD for retrieval.

¶Role

The official memory docs describe QMD as combining BM25, vectors, and reranking over local Markdown-backed memory. OpenClaw can opt into it with memory.backend = "qmd".

¶Scope

QMD is not OpenClaw’s canonical memory store. It is a retrieval layer that sits on top of files such as MEMORY.md and memory/YYYY-MM-DD.md to improve search and recall.

Using OpenClaw on Windows, March 2026

Mon, 09 Mar 2026 00:00:00 +0000

¶Purpose

This text collects practical advice for using OpenClaw on Windows as of March 9, 2026. The source base combines official OpenClaw documentation with two independent guides that contain operational advice rather than only product description.

¶Baseline: treat Windows as a WSL2 host

The official OpenClaw guidance is consistent: on Windows, run OpenClaw through WSL2, with Ubuntu recommended. The Windows page states that native Windows installs are untested and more problematic, while the CLI and Gateway are expected to run inside Linux.