Personal Model first

Elephant Agent starts from a person, not from a task. The Personal Model is the object that makes that possible.

It is the durable understanding Elephant Agent can carry into future help: stable enough to be useful, visible enough to correct, and small enough to avoid turning every transcript into truth.

The four lenses

Lens	The question it answers	Should change when...
Identity	Who is this person and how should Elephant Agent relate to them?	The user corrects names, language, role, boundaries, or stable preference.
World	What surrounds this person?	Projects, people, tools, locations, or operating context change.
Pulse	What is alive right now?	Priorities, pressure, energy, risks, or near-term focus shifts.
Journey	What has the path taught?	Decisions, lessons, repeated patterns, or long-running arcs become clear.

tip

When in doubt, ask whether a claim should help future Elephant Agent replies across multiple sessions. If not, it probably belongs in runtime history, current-turn recall support, or a task tool instead.

What belongs in the model

Candidate memory	Better owner	Reason
"User prefers concise answers."	Personal Model	Stable response style signal.
"The current build is failing on one test."	Episode/Step or task tool	Live work state, not durable identity.
"User might be switching projects."	Question	Uncertain signal that should be asked, not assumed.
"A search result supported the answer."	Current-turn recall support	Evidence for this turn, not truth by itself.

The correction contract

The Personal Model must remain correctable. Public and technical surfaces should make four moves feel natural:

Move	Meaning	Product behavior
Remember	Add a useful future-facing claim.	The claim can appear in future prompt projection.
Correct	Replace an outdated or wrong claim.	The old claim retires; the new one becomes active.
Forget	Retire a claim.	It should no longer shape replies.
Dispute	Keep uncertainty visible.	Elephant Agent avoids treating the claim as settled.

Not a hidden profile

The Personal Model is not a secret behavioral score. The user should be able to inspect claims, see support, answer or dismiss questions, and correct what does not fit.

Why capabilities orbit the model

Skills, tools, models, messaging, jobs, and embeddings all matter. But they do not replace the center.

Capability	What it adds	How it stays grounded
Skills	Repeatable workflows.	Skill affinity can be learned, but the user can inspect and toggle skills.
Tools	Action and retrieval.	Tool use records Steps and can update claims only through explicit paths.
Models	Dialogue and reasoning.	Provider choice changes how Elephant Agent thinks, not what is true.
Messaging	More surfaces.	The same elephant owns the continuity line.
Background jobs	Learning outside the immediate reply.	Jobs write through Personal Model tools, not a second hidden memory.