Ontology and Memory Systems (2/13) — Multi-Agent Ontology in Practice

OpenClaw 도구(Tools)와 훅(Hooks) 설계 — 최소 권한 원칙 적용기

Building a structure where multiple agents actually collaborate

Key Takeaways

  • Agent-to-agent communication requires two distinct paths: lightweight insight delivery and acknowledged task requests
  • Context isolation (1 agent = 1 role = minimum context) is the key to maintaining system quality at scale
  • Knowledge is managed across three layers: global CLAUDE.md → project CLAUDE.md → session memory
도구 프로파일과 그룹 제어

Background

Part 1 covered the concept and rationale of ontology — why defining relationships between agents matters, and what roles CLAUDE.md and memory serve.

Part 2 covers implementation. How agents exchange data, why context must be isolated, and how to layer memory.

Body

1. Message Hub — Designing Agent-to-Agent Communication

Once relationships are defined in the ontology, a channel for actual data exchange is needed.

Agent-to-agent communication splits into two types with fundamentally different characteristics.

① Insight delivery (fire-and-forget)

POST /api/webhooks
{"content": "...", "tags": ["analysis"], "source": "agent-name"}

A lightweight notification — "here's something I found." Send and done. No confirmation of receipt or processing.

Use cases: - An analysis agent forwards insights to a dashboard - A signal agent logs a state change

② Task request (ACK required)

POST /api/agent-messages
{
  "fromAgent": "home-server-manager",
  "toAgent": "write-score-director",
  "messageType": "task_request",
  "intent": "blog_material",
  "body": {"text": "Please write a blog post based on this material"}
}

This type requires an acknowledgment (ACK). It supports retry, thread tracking, and state management.

The receiving agent confirms and responds:

POST /api/agent-messages/{messageId}/ack
{"status": "acked", "response": {"text": "Received. Starting draft."}}

The reason for separating these two paths is clear. Requiring ACK on every message slows the system. Using fire-and-forget for everything drops critical tasks. Informational signals travel light; task handoffs travel with confirmation.

Practical example: The home server manager sends a blog material request to the write director via task_request. The write director ACKs and begins work. No human relay required — agents hand off work directly.

2. Context Isolation — Why Agents Must Be Separate

"Can't I just put every role into one powerful agent?"

This was tested. It does not work. Three reasons:

① Context contamination

When stock analysis context and Flutter UI code share the same conversation, the model conflates them. A prompt like "analyze the return value of this function" becomes ambiguous — stock data or code? This ambiguity emerges in practice.

② CLAUDE.md rule conflicts

Combining rules from multiple projects into a single CLAUDE.md creates contradictions. "Write concisely" (content) and "enumerate every edge case" (code) cannot coexist in the same instruction set.

③ Prompt length and performance degradation

Models tend to follow earlier instructions less faithfully as prompt length increases. Reducing irrelevant context is a performance optimization in itself.

Isolation principles: - 1 agent = 1 role = minimum context - When handing off a task to a sub-agent, do not pass the full session history - Pass only what the agent needs to know; isolate the rest

This principle is why quality holds even as the number of agents grows.

훅(Hooks) — 시스템 자동화의 축

3. Three-Layer Memory Architecture

Where and at what granularity to store knowledge matters. Three layers handle this.

Layer 1: Global CLAUDE.md — what every agent knows

~/.claude/CLAUDE.md  (symlinked across all agents)

Contents: - Full agent inventory and relationships (the ontology) - User judgment criteria and values - Inter-agent communication contracts - Cross-project rules

Every agent reads this at session start. Information here is shared across the entire system.

Layer 2: Project CLAUDE.md — what only this agent knows

{project-root}/CLAUDE.md

Contents: - Project purpose and structure - Project-specific rules and workflows - Skill triggers, agent role separation - Technology stack, API configuration

Project-specific knowledge is isolated so it does not contaminate other agents.

Layer 3: Session memory — what accumulates during conversations

{project}/.claude/projects/{id}/memory/

Contents: - User feedback ("don't do this," "that worked well") - Project state changes - Patterns learned mid-task

Persists across sessions. The agent for this project progressively calibrates to the user.

Flow between layers:

Repeated feedback in sessions
  → promoted to rule in project CLAUDE.md
    → if common across projects, promoted to global CLAUDE.md

Knowledge propagates upward. A principle validated in one agent eventually spreads to the full system.

4. Practical Tips for Ontology Management

Operational lessons collected over time.

① Keep a relationship table in CLAUDE.md

## Relationships
- OpenClaw → HomeServer: insights → dashboard display [data]
- QuietLeaf ↔ GrowNote: Flutter patterns shared [tech]
- MyLife → WriteScore: experience → content material [content]

Type tags ([data], [tech], [content]) make the relationship type readable at a glance.

② Checklist for adding a new agent

  • Add the role to the ontology
  • Define at least one relationship with an existing agent
  • If no relationship exists → reconsider whether the agent is necessary
  • Update global CLAUDE.md

③ Periodically validate relationships

Verify that relationships defined in the ontology are actually active. If the ontology says "A → B data transfer" but no transfer has ever occurred, that relationship must be removed or implemented.

④ Keep global CLAUDE.md concise

Ontology information, user preferences, communication contracts — that's it. Detailed implementation belongs in project CLAUDE.md. The shorter this file, the more reliably the model follows it, since it is read every session.

Design Considerations

  • A message hub is not required from day one. Human relay is sufficient early on — "take this agent's output and pass it to that one." Automate when the frequency justifies it.

  • The balance between isolation and sharing is critical. Over-isolation turns agents into islands; over-sharing contaminates context. Sharing belongs in global CLAUDE.md; isolation belongs in project CLAUDE.md. This boundary requires conscious management.

  • The ontology is a living document. Agents are added and removed, relationships change, roles evolve. It is not a one-time artifact — it requires continuous updates.

  • Replace agents incrementally. When replacing a stable agent, use parallel operation (Strangler Fig pattern) and a validation phase. If the new agent exhibits token runaway or unexpected behavior, the system must be able to revert to the original agent immediately.

Closing

One agent is a tool. Many agents without an ontology is chaos. Many agents with an ontology is a system.

The difference is not the number of agents — it is the design of relationships and the layering of knowledge. Share CLAUDE.md across agents, communicate through a message hub, isolate context, and layer memory — and the system maintains quality as it scales.

If you are adding agents one by one, now is the time to formalize the ontology.

댓글

댓글 쓰기

이 블로그의 인기 게시물

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

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 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"

이미지
← 8/9 Deep Learning Architec… 📚 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"

이미지
Series overview: Series index 📚 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. 핵심 요약 An evaluation set for RAG is ...
Read more »

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

이미지
← 7/9 Deep Learning Training 📚 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. 핵심 요약 CNN : weight sharin...
Read more »

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

← 6/9 Neural Networks 📚 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. 핵심 요약 SGD : \(w \leftarro...
Read more »

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

← 1/13 Current Structure Inve… 📚 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 📚 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 »