Skip to content

personal-ontology.md

Latest commit

 

History

History
107 lines (75 loc) · 3.82 KB

personal-ontology.md

File metadata and controls

107 lines (75 loc) · 3.82 KB
Raw

Personal Ontology

Durable AI memory that persists across conversations.

What this is for

Every conversation with an AI agent starts blank. By the end of the session it knows your name, your projects, your preferences — and then it forgets everything. The next session starts blank again.

A personal ontology is a persistent graph that the agent reads at the start of every session. It stores the things worth remembering: relationships, decisions, beliefs, projects, events, concepts. The agent picks up where it left off.

What to capture

Capture:

  • People you know and how you know them
  • Projects you're working on and their status
  • Decisions you've made and why
  • Events that happened (or will happen) and who was involved
  • Concepts you think about and how they connect
  • Places that matter and their relationships to people/events

Don't capture:

  • Ephemeral tasks ("send email to Alice today")
  • Drafts and work-in-progress that will be superseded
  • Information that changes faster than you'll query it
  • Noise — things you noted but never need to retrieve

The test: "Would I want an agent to know this two years from now?" If yes, it belongs in the graph. If no, it doesn't.

Start with the PersonalKnowledge template

start_here({ "template": "PersonalKnowledge" })

This seeds five classes purpose-built for personal use:

  • Person — people you know
  • Concept — ideas, domains, topics you think about
  • Event — things that happened or will happen
  • Location — places that matter
  • Document — notes, articles, books, conversations

Five classes is enough to start. Resist adding more until you have 20+ entities and natural clusters emerge that need their own class.

Key relations

Relation Domain → Range Use for
KNOWS Person → Person People who know each other
RELATED_TO Concept → Concept Concepts that connect
PARTICIPATED_IN Person → Event Who was at what
OCCURRED_AT Event → Location Where things happened
LOCATED_IN Person/Org → Location Where people/orgs are based
AUTHORED Person → Document Who wrote what
CITES Document → Document Reference chains

Seeding strategy

Start from sources you already have:

  1. Contact list — import key people with name and relationship context
  2. Calendar/Granola transcripts — events you participated in, decisions made
  3. Notes / journal — concepts you've been thinking about
  4. Projects — active work with dependencies and owners

Don't try to import everything. Import the 20 things you'd want an agent to know about you on day one.

Example session

// Start of session
health()          → {"status": "ok"}
start_here()      → {initialized: true, class_count: 5}

// Agent loads context
find_entities("Person", {})          → [Alice, Bob, Carol, ...]
find_entities("Project", {})         → [SparrowOntology, ...]

// Agent writes new information from the session
create_entity("Event", {
  "name": "Arch review with Alice",
  "date": "2026-03-23"
})
create_relationship(alice_id, event_id, "PARTICIPATED_IN")

The graph accumulates. Every session adds to it. Over time the agent has richer context than any single conversation could build.

Design principles

  • Classes are nouns. Person, Event, Concept — not PersonKnowledgeItem.
  • Don't create a class for what should be a property. "Alice is a founder" → Person.role = "founder". Not a Founder class.
  • If you can't name the relationship, you don't understand the connection. RELATED_TO is a placeholder. Replace it with something specific when you know what the relationship actually is.
  • 5 classes, not 20. Every class you add is one more thing to keep seeded and consistent. Start small.