Build assistants that truly know their users
Personal agents need to remember preferences, track evolving interests, and connect the dots between different aspects of a user’s life. With Headkey, your assistant builds a persistent model of who they’re helping.
Three Primitives, One Cognitive Architecture
Each primitive serves a different purpose. Here's how they work for this use case.
Memories
Remember everything the user shares
Preferences, past conversations, important dates, project notes — all searchable by natural language across every session.
{
"content": "User prefers dark mode in all IDEs and uses Vim keybindings.",
"intent": "observation",
"tags": [
"preferences",
"development"
]
}Beliefs
Build an evolving user model
"User prefers morning meetings" starts as a weak belief and strengthens over time. When preferences change, the old belief is superseded — not just overwritten.
{
"content": "User prefers morning meetings before 10am.",
"intent": "decision",
"confidence": 0.75
}Relationships
Connect people, projects, and interests
The agent maps how contacts relate to projects, how interests connect, and which topics come up together. Context is linked, not flat.
{
"content": "Sarah Chen leads the Q4 Product Launch.",
"intent": "relationship"
}Flat Memory vs. Structured Cognition
What changes when your agent has a mind, not just a vector store.
| Dimension | Flat Memory (RAG) | Headkey |
|---|---|---|
| Personalization depth | Keyword matching on stored notes | Evolving belief model with confidence scores |
| Preference changes | Old and new preferences conflict silently | Belief supersession tracks how preferences evolve |
| Connected context | Flat list of disconnected facts | Graph of people, projects, and interests |
| Privacy control | All-or-nothing sharing | Visibility scopes: private, scoped, or org-wide |
See It in Action
A personal assistant that learns preferences, builds an evolving user model, and uses intent: "correction" when preferences change.
|Start building your personal agent
Free to start. Plug-and-play cognition for any MCP-compatible agent in 60 seconds.