# Authority when pillars meet

How the Authoritative and Experiential pillars interact - reconciliation, uncertainty, and resolves_to.

SurrealDB Agent Memory’s memory model is built around **eight pillars** - not a flat “two layers” split. Two pillars matter most for day-to-day behaviour:

- **[Authoritative](/docs/agent-memory/architecture/eight-pillars-and-categories.md)** - curated artefacts: manuals, policies, product data, structured uploads. Higher default **trust**.
- **[Experiential](/docs/agent-memory/architecture/eight-pillars-and-categories.md)** - conversation and everything derived from it (reflection, elaboration, consolidation, …), including identity, knowledge, context, instructions, and uncertainties.

**Reconciliation** merges both streams with the same rules: supersession, **`uncertainty`** records, and trust logic apply whether facts arrived from a document or a turn. Physically, both are records in the **same** [unified substrate](/docs/agent-memory/mental-model/two-layer-architecture.md).

## Curated truth is never silently overwritten

**Authoritative** content does not change because a user said something different. When an experiential assertion conflicts with curated knowledge, SurrealDB Agent Memory:

1. Keeps the curated record unchanged.
2. Stores the user’s belief in experiential memory with provenance intact.
3. Surfaces the clash - for example via **`uncertainty`** records and conflict indicators in **`/state`** or **`/profile`**.

The user’s assertion remains valuable conversational context; it does **not** become canonical truth. That prevents incorrect chat from silently rewriting product, policy, or compliance data.

## Worked example

Your knowledge base states a **30-day return window**. A customer says: “Your return policy is 60 days.”

SurrealDB Agent Memory stores the customer’s belief under their scope in experiential memory and flags the disagreement. The curated policy record is unchanged. When the agent answers, it can acknowledge what the customer said while citing the policy that actually applies.

## Attribute resolution order

When both curated and experiential values exist for the same topic:

1. **Authoritative** values ground factual answers (price, policy, specification).
2. **Experiential** values personalise the interaction; conflicts are visible, not hidden.

Agents should **trust curated answers for compliance** and use experiential memory for preferences, history, and stated beliefs.

### Preferences over time

Two statements from the **same user** about the same topic - “always use formal tone” later replaced by “prefer casual” - are treated as **supersession**, not a contradiction. The older instruction stays in history with a bounded validity interval; the newer one is current. The agent behaves on what is valid **now**, and you can still ask what the preference was in March and when it changed. See [Supersession, decay, and forget](/docs/agent-memory/mental-model/memory-lifecycle.md).

## Linking chat to catalogue facts

Experiential mentions can link to authoritative **knowledge** records (for example resolving “AirPods Pro” in chat to the same product as your catalogue) via **`resolves_to`**. See [Cross-layer linking](/docs/agent-memory/reasoning/cross-layer-linking.md).

## Unreliable or satirical uploads

**Authoritative** means *curated by you for retrieval* - not *guaranteed true*. When you upload humour, parody, or deliberately false material (for example a satirical encyclopedia article), SurrealDB Agent Memory still indexes and retrieves it like any other document. At synthesis time the response model can recognise obvious parody and **refuse to treat it as factual**, especially when you ask what the **documents** say versus a general-knowledge question.

`source.kind = "document"` carries higher default **trust** for reconciliation against experiential chat - it does not disable the model's judgement at answer time. For compliance-sensitive deployments, curate uploads and use scope/labels so agents only retrieve vetted corpora.

## Practical guidance

- **Facts with regulatory weight** - ground in **Authoritative** content.
- **Personalisation** - lives in the **Experiential** pillar.
- **Contradictions** - surface them honestly; never pick silently.
- **Scope** - catalogue updates propagate across the Context; experiential records stay scoped per user, team, or project as you configure.

The full pillar list is in [Eight pillars and six categories](/docs/agent-memory/architecture/eight-pillars-and-categories.md).