Knowledge infrastructure
for AI agents

The database that can organize trillions of documents, memories and knowledge, map how it connects, and hand it to agents so they deliver high quality responses.


Fast and scalable agent graphs for the most accurate context and memory, built on object storage: up to 10-100x lower costs.

Trusted by teams building world-class agents

  • smallest.ai
  • composio
  • clinsight
  • Distyl
  • Genpact
  • Pocket
  • toptal
  • hsv.digital
  • gio
  • confida
  • klubi

Graph-based search for AI context. Build context and memory that continuously improve.

We’re a fast graph-based search engine that lets your AI explore all context so that you don’t have to go back and forth. Agents need context, memory, and a way to organize all of it

You build the layer for context. We give all the primitives:

  1. 01Assisted ontology using our model Typhon-1
  2. 02Minion-1 to extract, dream, and store context to continuously improve.
  3. 03Context with relations via API or MCP
  4. 04A unified graph to materialize views over any data source

Complete architecture

User

Context understanding

routing · entity extraction · query rewrite · safety flags

Retrieval orchestrator

Cypher reads / writes

Vectorstore (plugin)

semantic + bm25 + rerank

Connectors (plugin)

100+ sources: workspace, email, CRM

DB filters (plugin)

SQL / NoSQL

The core database

Fast, multi-tenant, and built on object storage

NAMESPACE multi-tenant isolation, auto-scalable

Writer node

Cypher writes → WAL + value log (S3 + disk cache)

Indexer node

reads WAL entries → builds index (GraphBLAS)

Reader node

Cypher queries → index + pending WAL = strongly consistent

Unified storage

S3 · object-storage native

namespace/data/wal

namespace/data/value

namespace/index

namespace/manifest.json

Tiered · Context flows across tiers on demand

Hot

in-memory cache

→

Warm

NVMe SSD

→

Cold

object storage

All primitives to build your own context and memory the way your agents want. Decisions, Traces, Docs, Relationships, Memories, Knowledge. With temporal awareness for when things evolve. What changed, When, Why. Give agents connected context with relations via API or MCP, and materialize views over any data source.

Services from functionSemanticKeyword “auth”
Status page, public, no auth
Admin routes reuse requireAuth for sessions
Checkout API calls requireAuth on every request
Reviewer: Payments teamPayments team
Ask

“What breaks if I change the auth middleware, and who should review it?”

  1. 01Traverse

    From the function to every service that calls it

  2. 02Search

    Semantic and keyword search over those services

  3. 03Resolve

    Ownership moved on 2 Nov, so the current reviewer wins

AnswerAll three steps run in one HydraDB query
Checkout API and Admin routes break. Platform team reviews

Agents become SOTA with graphs

We tested retrieval for various kinds of agents with HydraDB.

At scale, (vector) similarity isn’t relevance & Postgres isn’t enough

Similarity alone misses how context connects. Use a unified graph to materialize views over any data source, and give agents context with relations via API or MCP.

Comparison

Vector databases

SQL databases

Without graphs

With HydraDB

Similar ≠ relevant

Returns what’s close, not what’s related

Relevant, stateful context

Context that compounds as agents learn

Missed relationships

Concepts, entities and events stay disconnected

A full structured view

Every entity and link in one graph

Forgotten sessions

Traces and preferences lost between runs

Personalized recall

Results shaped by what agents have learned

Four stitched systems

VectorDB, GraphDB, Postgres and filesystems

One unified layer

Every primitive agents need, in one place

Recall degradation is the bottleneck

Embeddings have a theoretical ceiling. They compress each document into a fixed-size vector, limiting which combinations of relevant documents they can retrieve. Simply put: when multiple projects and owners appear across your data, finding similar passages may still leave your agent unable to reliably answer a precise question like “Who owns this project?”

Accuracy vs context length

ACCURACY %

100

80

60

40

20

0

HydraDB

90.79%

71%

38%

8K

32K

64K

96K

115K

CONVERSATION TOKENS

HydraDBVectorDBFull context

Open Source

HydraDB will be free forever

Free

Run open-source HydraDB on your own infrastructure

Self-host with any object storage compatible store

Import graphs from other databases

Community support

Ship

Run production on HydraDB

managed cloud

$25/mo

min usage

Fully managed by HydraDB

One database, unlimited collections

Assisted graph creation:

$0.70 per 1M tokens

Storage: $0.50/GB-month

100+ app connectors for easy start

Scale

Scale to thousands of customers with dedicated infrastructure

$300/mo

min usage

Assisted graph creation and cheapest storage for your graphs, plus:

Multi-tenancy: One database per customer

Private Slack channel

Advanced logs & monitoring

Response SLAs

Enterprise

Deployment and support tailored to your organization

Custom

Deploy in your cloud, or a dedicated cloud we manage

Dedicated forward-deployed engineer

99.5% uptime SLA

HIPAA BAA

Built by the community 13k Stars on GitHub & Counting

Frequently asked questions

HydraDB is a graph-based search engine for AI context, built on object storage. Use Typhon-1 for assisted ontology, and Minion-1 to extract, dream, and store context to continuously improve. Give agents context with relations via API or MCP.

HydraDB is for engineers building agents that need context, memory, and a way to organize it. Materialize views over any data source in a unified graph, so agents can explore documents, decisions, memories, and the relations between them.

You can. Postgres with a graph extension is a practical starting point. As your graph and query volume grow, graph indexes and traversals can compete with your core database for memory and CPU, slowing everyday transactions. HydraDB runs the graph workload on separate, purpose-built infrastructure, so you can scale graph queries without consuming the resources your application depends on

Use a vector database when you mainly need to find similar documents or passages. Use HydraDB when your application needs to follow relationships, traces, and context across multiple places: who owns a project, what depends on it, or which past decisions explain its current state. We don’t ask you to replace your vector database. Both graph search and vector search serve different purposes. In fact, HydraDB comes with built-in native support for vector search

Yes. HydraDB provides graph-based search for connected AI context. Use Typhon-1 for assisted ontology and Minion-1 to extract, dream, and store context to continuously improve, then retrieve context with relations via API or MCP for your GraphRAG application.