Semantic Memory

A living memory of
how your business runs.

A versioned knowledge graph of your metrics, definitions, decisions, and SOPs. Updates itself as your business changes. Every version preserved.

How it works.

  1. 01Butter ingests structured sources (dbt, Sheets) and unstructured sources (Notion, Slack, Confluence) into one graph.
  2. 02Each entity gets typed, linked, and time-bounded. Metrics, experiments, decisions, SOPs, teams - all in one graph.
  3. 03Every change is versioned. Conflicts surface for review. Roll back any definition. See what the AI knew at any point in time.

Semantic Memory vs. dbt semantic layer.

dbt semantic layer / CubeButter
Who owns itData engineers, via YAML and DAG knowledgeAnalysts and business users, no YAML required
What it capturesWhat data is, schema-shapedWhat data means, plus decisions, SOPs, experiments
How it changesManual edits, PRs, releasesLearns from usage, surfaces conflicts, updates with approval
HistoryGit history of YAML filesVersioned graph of every definition, decision, and conflict

Questions.

Is Semantic Memory a knowledge graph?

Yes - and a living one. It connects metrics, definitions, decisions, and SOPs as typed, linked entities. Every change is versioned; conflicts surface for review instead of definitions silently drifting across tools.

How is this different from dbt semantic layer or Cube?

Those tools were built by data engineers, for data engineers. They formalize what data is, but not what it means to the people running the business. They also stop at the data layer. Semantic Memory captures business meaning, decisions, and SOPs alongside the data, owned by the analysts who use it.

Can I see the full history of a definition?

Yes. Every edge in the graph has a commit time and a valid-from / valid-to. You can answer 'what was our definition of GMV before the November revision?' in one query, and trace any AI answer back to the exact version of context it used.

How does it stay accurate?

Butter learns from usage. When a definition is queried often, it's reinforced. When a new piece of context conflicts with what Butter already knows, Butter pauses, surfaces the conflict, and asks the owner to resolve it with a few targeted yes/no questions. Every resolution becomes a new versioned edge.

What sources does it ingest?

Notion, Confluence, Google Sheets, Slack, dbt, Looker semantic layer, Cube, and any source exposed via MCP. New connectors ship monthly. Butter doesn't replace these systems. It reads from them and keeps a unified, versioned memory across them.

Is this a replacement for our data warehouse?

No. The warehouse is your source of truth for raw data. Semantic Memory is your source of truth for what that data means, who owns it, how it has changed, and what the business decided about it.

Stop maintaining a static layer. Have a knowledge graph that maintains itself.

We are rolling out access by team. Tell us where stale definitions are slowing your analysts down. We'll reach out within a few business days.

Request access