Context Lake
The Context Lake is Zep’s system for managing and serving Context Graphs. It unifies context from business data, documents, and conversations while preserving the scope of each graph and the access controls that apply to it.
The service manages millions of Context Graphs on Konig, Zep’s graph database service. This count describes the empirical scale of the Zep service. The count is not a per-account capacity limit or an availability guarantee.
One system, many Context Graphs
Each Context Graph represents facts, relationships, and changes over time for a subject. A subject can be a user, customer, account, project, product, or business domain.
Applications can use:
- A user graph for context that belongs to one application user.
- A Context Graph addressed with
graph_idfor shared customer, account, organization, or domain context. - Multiple graphs when a task requires both user-specific and shared context.
The graph scope remains explicit. Shared context is available only inside the project and policy boundaries that authorize it.
The context lifecycle
Zep applies one lifecycle across the three solution categories:
- Ingest: Add business data, documents, JSON, or conversations.
- Shape the graph: Define the entities, relationships, instructions, and summaries that are important for the domain.
- Retrieve: Search a graph or assemble a Context Block for a user and thread.
- Govern: Control access, trace facts to source episodes, and review dashboard or API activity.
Start with Working with Context for the implementation lifecycle.
Solutions built on the Context Lake
Combine business data and documents into shared context for agents across an organization.
Provide context from previous conversations, user activity, and changing preferences.
Combine records, conversations, and events so agents can understand a customer or account.
Temporal context and source traceability
When new information contradicts an existing fact, Zep can invalidate the old fact and keep its validity history. An application can retrieve what is true now or search graph data for a past time.
Source episodes provide provenance for facts and graph artifacts. This provenance lets an application trace retrieved graph data to the source data from which Zep derived it. The application must retain its retrieval references if it must connect a generated answer to those sources.
Flex Plus and Enterprise accounts can also use Observations. These records describe patterns that Zep identifies in graph data.
Governance and application responsibility
Governance includes team access, policy-based access to context, source traceability, Audit Logs, and API Logs. API Logs support query-audit workflows within the request information that they record.
Context supports correct task completion, but it does not guarantee model behavior or answer accuracy. Zep access policies control access to Zep context. Your application remains responsible for tool permissions, action authorization, model instructions, and validation of external actions.