What if your AI agent truly knew you — your goals, your communication style, your preferences — and adapted every response accordingly? That’s what Personalized Agents is about: building AI agents that carry a genuine understanding of you, powered by the OpenClaw framework.

This is an active project. The full workspace configuration and custom skills live at github.com/diepdaocs/agents.

The Problem

Most AI tools treat every conversation as a blank slate. You re-explain your context, re-state your preferences, and get generic responses that don’t account for who you are or what you’re trying to accomplish.

A personalized agent fixes that by maintaining a persistent profile — your identity, your goals, your working style — and weaving that into every interaction.

Architecture

Personalized Agents Architecture

Layers at a glance

User Layer — you define who you are through workspace markdown files. The Profile Builder compiles these into the agent’s operating context.

OpenClaw Gateway — the central service that discovers skills, routes requests, manages auth, and connects to your chosen LLM (Claude, GPT-4o, Gemini, or a local Ollama model).

Workspace Layer — the workspace/ directory is the brain. Each markdown file shapes a different aspect of agent behaviour.

Skill Agents — auto-discovered from workspace/skills/, each skill adds a specialized capability: memory recall, context tracking, task execution, persona shaping, or scheduled heartbeats.

Interface Layer — interact via terminal TUI, single-turn CLI commands, or the HTTP gateway API from your own applications.

Workspace Files — Giving the Agent a Self

File What it controls
SOUL.md Agent personality, tone, and values
IDENTITY.md Who the agent is and how it introduces itself
USER.md Your preferences, background, and working style
AGENTS.md Operating rules and behavioral constraints
TOOLS.md What tools and integrations are available
HEARTBEAT.md Scheduled tasks and proactive reminders
MEMORY.md Long-term memory across sessions

These files are plain Markdown — edit them like a README and the agent updates immediately.

Step-by-Step: Building Your Personalized Agent

Step 1 — Install OpenClaw

pip install openclaw

Verify the installation:

openclaw --version
openclaw doctor

openclaw doctor checks your environment: model connectivity, workspace structure, and skill discovery health.

Step 2 — Clone the Workspace

git clone https://github.com/diepdaocs/agents.git
cd agents

The workspace/ directory is the core of your agent configuration. Everything else is commentary.

Step 3 — Connect Your LLM

Set your API key as an environment variable (pick your provider):

# Anthropic Claude
export ANTHROPIC_API_KEY=sk-ant-...

# OpenAI
export OPENAI_API_KEY=sk-...

# Google Gemini
export GOOGLE_API_KEY=...

# Local Ollama (no key needed)
ollama pull llama3

OpenClaw auto-detects which key is present and picks the model accordingly. You can also fix a model in config/settings.json:

{
  "model": "claude-3-5-sonnet-20241022",
  "workspace": "./workspace"
}

Step 4 — Define Your Profile

Open workspace/USER.md and describe yourself. The more specific, the more the agent adapts:

# About Me

- **Role**: Software engineer with a focus on backend systems and data pipelines
- **Communication style**: Direct and concise. Skip preamble.
- **Tech stack**: Python, Go, ClickHouse, Kafka, Redis
- **Working hours**: SGT (UTC+8), mornings are focus time
- **Current goals**: Ship Personalized Agents v1 by end of Q1 2026

Then edit workspace/SOUL.md to shape the agent’s personality:

# Soul

You are a pragmatic technical collaborator. You prefer concrete examples over
abstract explanations. You call out trade-offs honestly. You celebrate shipped
code, not theoretical perfection.

Step 5 — Launch the Gateway

Start the OpenClaw agent service:

openclaw agent serve --workspace ./workspace

The gateway:

  • Scans workspace/skills/ and auto-discovers all skill directories
  • Loads your profile from the workspace markdown files
  • Starts an HTTP server (default: localhost:8080)
  • Connects to your configured LLM

Check which skills were discovered:

openclaw skills list

Step 6 — Run Your First Personalized Interaction

Single-turn command:

openclaw agent run "Summarise what I should focus on today"

The agent draws on your USER.md goals, HEARTBEAT.md schedule, and MEMORY.md history to give you a response tuned to your context — not a generic answer.

Interactive TUI:

openclaw tui

The terminal UI gives you a chat-like experience with full conversation history, session memory, and inline skill invocations.

Step 7 — Add a Custom Skill

Skills live in workspace/skills/<skill-name>/. Create a new one:

mkdir -p workspace/skills/standup-prep

Add a skill definition file workspace/skills/standup-prep/skill.md:

# Standup Prep

Generates a daily standup summary by reviewing recent MEMORY.md entries and
current goals from USER.md.

## When to invoke
When the user asks "what did I do yesterday", "prepare my standup", or
"standup summary".

## Output format
- **Yesterday**: bullet list of completed items from memory
- **Today**: top 3 priorities from USER.md goals
- **Blockers**: any flagged items from AGENTS.md constraints

Restart the gateway — the skill is auto-discovered, no install or registration needed:

openclaw agent serve --workspace ./workspace

Invoke it:

openclaw agent run "prepare my standup"

Step 8 — Set Up Scheduled Heartbeats

Edit workspace/HEARTBEAT.md to add proactive behaviours:

# Heartbeat Rules

## Morning (09:00 SGT, weekdays)
- Pull today's calendar context
- Surface top 3 goals from USER.md
- Remind of any overdue items in MEMORY.md

## Weekly (Friday 17:00 SGT)
- Generate a week-in-review from MEMORY.md
- Suggest what to carry forward vs. close out

The heartbeat agent reads these rules and fires them on schedule via the gateway’s background scheduler.

Step 9 — Verify and Diagnose

At any point, run diagnostics:

openclaw doctor

This checks:

  • Workspace file validity
  • Skill discovery status
  • LLM connectivity and latency
  • Memory store integrity

Custom Skills Included in the Repo

The diepdaocs/agents repository ships with two example skills to get you started:

daily-briefing

Generates a morning briefing by combining:

  • Your goals from USER.md
  • Scheduled items from HEARTBEAT.md
  • Recent memory entries from MEMORY.md

Invoke with: "Give me my daily briefing" or "Morning summary"

file-organizer

Scans a target directory and proposes an organization structure based on your preferences in USER.md. Uses the LLM to infer sensible groupings without hard-coded rules.

Invoke with: "Organize my downloads folder" or "Clean up ~/projects"

What Makes It Personal

The magic is in the composition. Every agent response is shaped by:

  1. Your profile (USER.md) — the agent knows your role, stack, and preferences
  2. Your persona definition (SOUL.md) — the agent speaks in a style that resonates with you
  3. Your history (MEMORY.md) — the agent remembers what you’ve worked on across sessions
  4. Your schedule (HEARTBEAT.md) — the agent is proactive, not just reactive
  5. Your rules (AGENTS.md) — the agent respects your constraints and operating norms

Swap out the workspace files and you have a completely different agent — same framework, different person.

Technology Stack

Component Technology Role
Agent Framework OpenClaw Orchestration, skill discovery, gateway
LLM Claude / GPT-4o / Gemini / Ollama Language understanding and generation
Profile Store Markdown files (workspace/) Human-readable agent configuration
Skill Runtime Python Custom skill implementation
Interface CLI / TUI / HTTP API Multiple interaction modes
Scheduler OpenClaw Heartbeat Scheduled and proactive tasks

Repository

Source code, workspace files, and skill examples: github.com/diepdaocs/agents

Clone it, fill in your profile, and you’ll have a personalized agent running in under 10 minutes.

Updated:

Comments