OpenClaw Build and Operations (1/5) — OpenClaw Anatomy: Components of a Single-User Agent

Pre-Migration Inventory Framework — Agents · Cron · LaunchAgent · Scripts · Skills · Memory Tiers

ํ•ต์‹ฌ ์š”์•ฝ

  • OpenClaw is a single-user agent system composed of multiple agents, schedulers/daemons, operational scripts, and skills.
  • Memory follows a 4-tier structure (Identity / Curated / Project / Episode), queried via oMLX bge-m3-mlx-fp16 embeddings with hybrid 0.7/0.3 + MMR + 30-day temporal decay.
  • This document defines an inventory framework for pre-migration or pre-redesign analysis, structured across seven layers: agent / scheduler / OS daemon / script groups / skills / memory tiers / routing.

Why Inventory Comes First

When migrating or redesigning an agent system into a different framework, the first task is not code porting — it is building a component inventory. Three reasons:

  1. Rhythm lives in the scheduler, not in the code. Omitting the cron/LaunchAgent list erases the rationale for why individual scripts exist.
  2. Persona count ≠ agent count. The number of agents reflects model routing and tool-permission splits, not the number of conversational interfaces. Without this distinction, the target system ends up with an over-engineered profile count.
  3. Memory tiers have different portability units. Identity / Curated / Project / Episode each carry distinct storage backends, update cycles, and write-authority rules. Tier-by-tier mapping is required to prevent data loss.

The following uses OpenClaw as a reference example. The same framework applies to any multi-agent system.

1. Agent Layer — One Conversational Interface + N Backend Workers

The only conversational interface is mir; all other agents are backend workers. Agents are split along two axes: model tier and tool permissions.

Representative split:

Role Group Model Tier (example) Responsibility
Conversational interface High-judgment LLM Sole user-facing entry point, routing
System monitor Local oMLX System state tracking
Research Lightweight LLM Web research, AutoSearch Deep
Writing Mid-tier LLM Content generation
Coding / orchestration Code-specialized LLM CLI, sub-agent / skill factory
Memory semantic judgment Mid-tier LLM Reflect decisions
Memory I/O Local oMLX bank/ file I/O
External integrations Lightweight LLM Google Workspace, messenger delivery, etc.

(Plus additional backend workers)

How it works: The persona exposed to the user is a single entity (mir). Behind it, tasks are distributed based on model tier, cost, and latency characteristics — local oMLX/Gemma for embedding and file I/O, gemini-2.5-flash for semantic judgment, gpt-5.4 high for high-judgment tasks.

2. Scheduler Layer — System Rhythm via Cron

Representative tasks by time resolution:

Frequency Task Type
Daily (early morning) Reflect — semantic consolidation
Hourly Memory micro-cycle — short-term memory cleanup
Daily (night) Self-review — behavior audit, daily-summary
Minute-level polling External source scan (research, communities)
Monthly Long-term memory review

(Plus additional scheduled tasks)

How it works: Cron is not a simple scheduler — it defines the agent's cognitive rhythm. Morning Reflect (semantic consolidation), hourly micro-cycle (short-term cleanup), nightly self-review (behavior audit), and monthly review (long-term pattern compression) each run at distinct time resolutions to prevent memory volatility. The migration question is: "which target system can preserve these four temporal cycles?"

3. OS Daemon Layer — macOS LaunchAgent

Processes that must persist at the OS level:

  • ai.openclaw.gateway — Gateway daemon (port 18789)
  • ai.openclaw.update-check — Auto-update daemon
  • com.openclaw.omlx / omlx-proxy — Local embedding model server
  • com.openclaw.daily-audit — 05:00 daily audit Phase A–E
  • com.claude-agent.watchdog / autostart — Session watchdog

Constraint: LaunchAgent is macOS-specific. Porting to Linux or a container target requires 1:1 replacement with systemd units or supervisord. Port binding (18789) and audit Phases A–E can be decoupled into an OS-independent layer.

4. Script Layer — Operational Scripts, Grouped by Function

Scripts belong to the following functional groups:

  • Memory pipeline: retain-extract/merge, recall-tree/match, confidence-decay, topics-validate/expand, user-pattern-stage.
  • Heartbeat: heartbeat-router/update/tick/health, proactive-exception-alerts.
  • Self-review: self-review-prescan/apply/fix, skill-candidate-stage.
  • Daily audit: docs-snapshot, release-check, agent-linter, config-drift, cron-analytics, session-audit.
  • External integrations: scrape.py (Scrapling + playwright), gws-wrapper, omlx-serve.

Portable patterns: confidence-decay and topics-validate/expand from the memory pipeline are self-contained units portable to any long-term memory system. The Heartbeat group approximates a standard pattern for distributed agent health monitoring.

5. Skill Layer — General-Purpose + System-Specific

  • General-purpose workflows: brainstorming, writing-plans, verification, deep-interview, code-review, testing, git-commit, project-doctor, self-audit.
  • OpenClaw-specific: oc-config, oc-agent, oc-channel, oc-skill, oc-memory, oc-doctor, oc-deploy, oc-backup.

Separation criterion: General-purpose skills are framework-agnostic and can be carried over verbatim. System-specific skills are coupled to the configuration schema and deployment paths of OpenClaw; they require rewriting against equivalent concepts in the target system.

6. Memory v0.6 — 4-Tier Structure

  • Identity: MEMORY.md (~600 tokens). Fixed identity and preferences. Lowest change frequency.
  • Curated: bank/ (world / experience / opinions / patterns / topics / index). Synthesized knowledge. Updated by Reflect.
  • Project: bank/entities/. Per-entity accumulation.
  • Episode: memory/ (daily logs). Raw source events.

Retrieval mechanism: oMLX bge-m3-mlx-fp16 embeddings + hybrid 0.7/0.3 (vector / keyword) + MMR + 30-day temporal decay. Reflect runs as a 6-phase pipeline; U-tag processing (user-relevance judgment) follows a 3-phase dialectic: observation → hypothesis → verification.

Migration perspective: Each tier has distinct storage cycles and write authority. Identity is manually managed; Curated is updated by Reflect (the semantic-judgment agent); Project is mixed; Episode is auto-appended. If the target system does not support this authority split per tier, consistency breaks.

7. Channel Routing Layer

  • One primary Discord channel carries all Reflect outputs and self-review results.
  • Telegram Forum with multiple topics: General / Server / Bot / Research / Memory / Heartbeat.

Design rationale: Discord handles a single high-density stream of judgment outputs; Telegram handles classified real-time events across multiple topics. The split is governed by a trade-off between event frequency and reading cost.

Two Structural Facts the Inventory Reveals

Completing the floor plan surfaces two structural facts.

First, memory · heartbeat · self-review are rhythm, not code. Cron/LaunchAgent defines the system's cognitive cycle. Whether the target container can preserve that cycle is the core portability criterion.

Second, agent count is not persona count. Of all agents, only one is the conversational interface; the rest are workers partitioned by model tier and tool permissions. Therefore, a target system can achieve equivalent functionality with fewer profiles.

Applicability and Open Questions

Applicable to: - Multi-agent systems being ported or redesigned into a different framework. - Audit and migration checklists for systems with a 2-tier or deeper memory hierarchy. - Rhythm-preserving design for long-running agents based on cron/LaunchAgent.

Open questions: - Where is the optimal boundary between Curated and Project in a 4-tier memory system? In entity-centric systems, the two tiers may be mergeable. - What is the sustainable upper bound for a single-persona + N-worker model? As the agent count grows, routing-decision cost likely increases super-linearly. - Does Reflect quality (semantic consolidation) hold when cron-based rhythm is replaced with an event-driven model?

The next post presents a 1:1 mapping of the above layers against Hermes built-in capabilities, across four categories: Replace / Partial Replace / Port As-Is / Deprecate.

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »