API
Write
mutate
Read
search (BFS, PageRank, Semantic, BM25)·patterns·list / get
Typed surface for mutations, traversals, graph algorithms, and lookups.
The runtime underneath Zep
The Context Graph Engine
Most graph databases are built to hold one large graph. The Context Lake workload is the opposite: millions of smaller graphs, mostly cold, all temporal, all governed. Different workload, different runtime.
Architecture
Built for the workload, not retrofitted.
The Engine is shaped for the Context Lake workload from the data model up. Millions of graphs. Sparse activity per graph. High aggregate throughput. Governance applied independently to each.
Every architectural decision — tiered storage, in-memory adjacency, native ABAC, bi-temporal edges — follows from the workload.
WAL
append-only·replicated·ordered
Mutations durably appended to the write-ahead log before acknowledgment.
Graph Primitives
Algorithms
BFS·PageRank·pattern matching·path analysis·temporal weighting
Structure
adjacency & CSR matrices·vector index·BM25 index
Graph algorithms, traversals, and pattern matching run over compact in-memory structures.
Governance
ABAC·multi-tenant isolation·customer key encryption·retention policies·audit·provenance
Native to the substrate, not a layer bolted on. Every read and write is policy-gated for access and provenance; retention runs across the data lifecycle.
Tiered Data Layer
HotRAMmicroseconds
WarmLocal NVMemicroseconds
ColdObject Storemilliseconds
Hot graphs serve at memory speed. Inactive graphs are evicted to NVMe and object-store, and rehydrated in milliseconds.
Durable Storage
WAL·content·metadata·graphs·indexes
Snapshots and WAL persist on a multi-AZ, highly durable object and document store.
Scale
Scaling in constant time.
Retrieval latency holds near-constant as the graph count grows. Current production runs sustain thousands of mutations and queries per second across millions of graphs — p50 latency unchanged from a thousand graphs to a million.
Retrieval latency vs. graph countproduction · last 30 days
p50 retrievalp95 retrieval
Tiered storage
Storage that tracks activity.
Three tiers. Hot graphs live in RAM at microsecond latency. Warm graphs sit on local NVMe. Cold graphs rest on object storage and rehydrate in milliseconds.
Cost tracks active graphs, not total graphs. A deployment with one million graphs and one percent of them hot pays for one percent of the memory.
§ 05 · Retrieval
Unified retrieval.
Vector similarity, BM25, graph traversal, and pattern matching run over the same data. One query, one ranked answer — no separate retrieval stack to stitch together.
§ 06 · Primitives
Graph primitives in memory.
Hot graphs are held as adjacency lists and CSR matrices. Cache-friendly, deterministic layout, ready for matrix operations.
BFS, PageRank, pattern matching, path analysis, and temporal weighting run at microsecond latencies over those structures.
§ 07 · Governance & temporality
Governance and temporality, native.
ABAC, multi-tenant isolation, customer key encryption, retention policies, audit, and provenance are properties of the substrate, not a layer above it. Every read and write is policy-gated.
Every edge carries four timestamps. Point-in-time queries, automatic invalidation, and temporal weighting follow from the data model.
§ 08 · Durability
Durable by default.
Mutations are appended to a replicated, ordered write-ahead log before acknowledgment. Snapshots, indexes, and content persist on multi-AZ durable storage. Point-in-time recovery across the lifecycle.