Personal Ontology Skill

A framework for organizing your life as a knowledge graph. Objects are the entities (beliefs, goals, projects). Links are the relationships between them (serves, supports, contradicts). The agent uses this graph to make decisions aligned with who you are.

Quick Start (Example)

You tell the agent: "Moltbot bootstrap personal ontology." It scans your notes, proposes candidate Objects (e.g., a Belief about AI, a Goal for health, a Project for a newsletter), and presents them for review - nothing is auto-committed.

You confirm or edit those candidates. The agent then creates/updates your ontology files and links Projects → Goals → Core Self, flagging any orphans or contradictions for your decision.

From that point on, the agent runs a light daily scan: it watches for new beliefs, predictions, goals, and projects, and surfaces only high-confidence candidates or conflicts so you stay aligned without extra maintenance.

The Object Hierarchy

Objects are organized from most abstract/stable to most concrete/changeable:

1. 1. Higher Order - The highest organizing principle (God, universe, truth). Acknowledged, not defined.
2. Beliefs - Foundational assumptions about reality. What you hold to be true. Generally unfalsifiable.
3. Predictions - Your model of what will happen. Testable, time-bound, updateable.
4. Core Self - Who you are: Mission, Values, Strengths.
5. Goals - Time-bound objectives serving your Core Self. Outcomes you want to achieve.
6. Projects - Organized efforts toward goals. Bounded work with beginning and end.
7. Tasks - Atomic units of work. (Live elsewhere: daily notes, Kanban, reminders.)

Link Types

Every Object (except Higher Order) should link to other Objects. Standard link types:

Link	Meaning	Example
INLINECODE0	Directly supports an outcome	"This Project serves Goal X"
INLINECODE1

Provides evidence/foundation for | "This Prediction supports Belief Y" |
| contradicts | In tension with | "This Belief contradicts Prediction Z" |
| relates-to | General association | "This Goal relates-to Value W" |
| depends-on | Requires for completion | "Project A depends-on Project B" |
| evolved-from | Updated version of | "Prediction 2.0 evolved-from Prediction 1.0" |

Validation rule: Every Project must serve at least one Goal. Every Goal must serve Core Self. Orphan Objects are flagged for review.

File Structure

Live ontology (canonical): INLINECODE8

CODEBLOCK0

Each file contains multiple Objects of that type, each with a ## Links section.

Suggestions queue: INLINECODE10

Use this file to capture all candidate updates (bootstrap + ongoing).

---

For AI Agents

Deployment Modes

Interactive Mode: Direct conversation. User asks for help, agent references ontology for context.

• - "Should I take this job?" → Check against Goals, Values, Mission
• "What should I work on?" → Surface high-priority Projects serving active Goals

Embedded Mode: Agent uses ontology to inform all decisions without explicit reference.

• - Email triage → Prioritize based on Goals/Projects
• Task suggestions → Only suggest what serves active Projects
• Calendar optimization → Protect time for Goal-aligned work

Automated Mode: Passive scanning and maintenance without user prompting.

• - Daily scan for new/changed Objects
• Flag contradictions and orphans
• Surface stale Predictions

When to Reference the Ontology

1. 1. Making decisions - Check proposed action against Values, Goals, Mission
2. Prioritizing - Use Goal hierarchy to rank options
3. Explaining - Connect tasks to higher purpose ("This serves your goal of...")
4. Flagging misalignment - "This task doesn't connect to any active Goal"
5. Surfacing insights - "Your Prediction about X has a time horizon of Q1 - any update?"

Integration Recipes (optional)

For users who want the ontology to shape daily behavior, integrate it into:

• - Morning briefing: restate Mission + top Goal + active Project, then suggest the next concrete output.
• Task sanity checks: flag tasks that don't clearly map to a Project → Goal.
• End-of-day closure: reflect on what served the Mission/Goals, then close the loop so the user can disconnect.
• Wind-down: brain dump + tiny alignment cue for tomorrow.

Adaptive prompts ("magic in the background")

If higher layers are missing (Higher Order, Predictions), do not nag daily.
Instead:

• - Ask one tiny, optional prompt at most once every N days.
• Track last-asked timestamps in a small local state file (example: memory/ontology-nudges.json).
• Only surface prompts at natural transitions (work start or wind-down), and skip on overloaded days.

These can be implemented via scheduled prompts (cron) that read the ontology files + the user's task board/journal + the nudge state file.

How to Use It

CODEBLOCK1

---

Daily Passive Scan

Agents should perform a lightweight scan of recent activity to maintain the ontology.

What to Scan

• - New notes (last 24h) - Look for ontology-relevant content
• Modified files (last 24h) - Check if existing Objects need updates
• Task completions - Do they affect Project status?
• Calendar/journal - Any events that validate/invalidate Predictions?

Extraction Patterns

When scanning, look for these signals:

Pattern	Candidate Object
"I believe...", "I think...", "I'm convinced..."	Belief
"I predict...", "By 20XX...", "will happen", "I expect..."

Prediction |
| "My goal is...", "I want to achieve...", "I'm working toward..." | Goal |
| "I'm working on...", "Project:", "I'm building..." | Project |
| Mission statements, value lists, "What matters to me..." | Core Self |

When to Act

Signal	Action
New Object candidate (high confidence)	Queue for user review
New Object candidate (low confidence)

Note in daily memory, don't surface yet | | Existing Object contradicted | Surface immediately with evidence | | Prediction time horizon passed | Prompt for resolution | | Project completed | Prompt to update Goals |

Contradiction Detection

When new content conflicts with existing ontology:

1. 1. Note the specific contradiction
2. Surface to user with both sides
3. Don't auto-resolve - user decides which to update
4. Track resolution in Prediction Log or Object history

---

Intelligence Layer

The ontology isn't just storage - it drives insights. Regularly surface:

Orphan Detection

• - Orphan Project - Doesn't serve any Goal → "This project isn't connected to your goals. Is it still relevant?"
• Orphan Goal - Doesn't connect to Core Self → "What mission/value does this goal serve?"
• Orphan Task - Doesn't belong to any Project → "Is this task important enough to track?"

Staleness Checks

• - Stale Prediction - Time horizon passed, no update → "Your prediction that X would happen by Q1 2026 - did it?"
• Stale Project - No activity in 30+ days → "Is [Project] still active?"
• Stale Goal - No serving Projects → "What's the next project for [Goal]?"

Alignment Checks

• - Task audit - "These 5 tasks from today don't connect to any Project. Intentional?"
• Time allocation - "You spent 80% of this week on Goal 2 but marked Goal 1 as top priority"
• Value drift - "Your recent decisions seem to prioritize X over Y, but your values list Y first"

Pattern Recognition

• - Recurring themes - "You've mentioned 'AI safety' in 5 notes this month. Should this be a Belief or Goal?"
• Implicit Objects - "You act as if you believe X, but it's not in your Beliefs. Add it?"
• Prediction clusters - "These 3 predictions are related. What's the underlying model?"

---

Review Cadence

Weekly (Agent-initiated)

• - Are current tasks serving projects?
• Are projects serving goals?
• Any new Objects to add from the week's notes?

Monthly (User-prompted)

• - Do goals still serve Core Self?
• New predictions to add? Existing ones to update?
• Any completed Projects to close out?

Quarterly (Deep review)

• - What surprised you? What does that reveal?
• Has Core Self shifted?
• Full Prediction Log review - what were you right/wrong about?

---

Temporal Tracking

Objects evolve. Track when and why:

CODEBLOCK2

For Predictions specifically, track:

• - Created: When you made the prediction
• Timeframe: When you expected resolution
• Resolved: Date + outcome (confirmed/disconfirmed/modified)
• Learning: What the outcome revealed

---

Setup

How to Run Bootstrap (User-Facing)

Say: "Moltbot bootstrap personal ontology."

The agent will:
1) Scan your notes for candidate Objects
2) Present candidates for your review (no auto-commit)
3) Write/merge confirmed Objects into your ontology files

Default location (Obsidian): Vault v3/ontology/ (pretty-formatted, readable Markdown)

For New Users

1. 1. Run the bootstrap process (see bootstrap.md) to extract candidate Objects from existing notes
2. Review and confirm/edit candidates
3. Work through prompts.md to fill gaps
4. Agent begins daily passive scans

For Existing Users

1. 1. Copy templates to your ontology folder
2. Fill in what you know
3. Agent maintains and extends over time

Reference Implementation

See templates/ for starter files. The user's ontology will be created in their notes folder.

Files

• - SKILL.md - This file (agent instructions)
• INLINECODE16 - Rules for categorization and validation
• INLINECODE17 - Initial extraction from existing vault
• INLINECODE18 - Guided questions for building each layer
• INLINECODE19 - Starter files for each Object type