Skip to main content
The Context Graph is persistent communication memory. Every conversation makes the next one smarter: outcomes, preferences, policies, and history consolidate into one context layer that agents read before they speak. Most agent memory is a vector store in someone else’s cloud - opaque blobs retrieved by similarity, with no answer to “why does the agent believe this?” The Context Graph is built differently. Every fact carries its source, its timestamp, a validity window, and a sensitivity level. Retrieval respects tenant, permission, and purpose - not just similarity. A fact learned on a support call does not leak into a sales pitch, and a fact that expired last quarter stops being retrieved. What the graph holds: customer context, your knowledge base, the workflow model, the action graph, communication policy, the evaluation suite, outcome history, and optimization history. All of it is inspectable and exportable - export and erase APIs are part of the early-access surface, so the memory your agents run on is never a black box you cannot audit or delete.

​
The primitives

​
On the spine

Context underlies the spine from entity through outcome. The Context Graph reads entity records, conversation history, and structured outcomes, and writes back consolidated facts, relationships, and summaries. Agents and Flows read from it before every interaction; each interaction’s outcome feeds the next consolidation pass. Changes to context will surface on the unified event stream as it ships (see Events).

​
What stays hidden

The embedding and vector implementation is deliberately not exposed. Which models embed what, how the index is built, and how similarity is scored are internals we tune and swap freely. You work with facts, relationships, summaries, and retrieval - never with vectors.

​
Status

Early access. Provenance, validity windows, and sensitivity levels are the early-access core. Two pieces carry production traffic today under other names: post-call summaries via the Post-Call Data API, and knowledge base retrieval in the Studio knowledge base. The unified graph - one context layer across channels, with export and erase APIs - is what early access opens up.

​
Go deeper

Knowledge Base

Where knowledge retrieval lives in the Studio today.

Sessions

How conversation state works inside a live SuperDialog session.

Entities

The identities that context attaches to.

Intelligence

The blocks that extract and score what the graph consolidates.