HydraDB vs Zep

tl;dr

Zep is a managed memory and context service for agents. It decides how memories are stored, which kinds of facts exist and when old ones retire. HydraDB is the graph underneath decisions like those: you choose what memory and context mean for your agents, the structure evolves with your data, and the graph stays yours to query.

Based on public docs and pricing as of Nov 2026. Something out of date? Suggest a correction →

Zep gives agents memory and company context as a managed service. Conversations, documents and business data become a graph where every fact knows when it became true and when it stopped. Old facts are kept rather than deleted, and each one traces back to its source. Connectors bring in Slack, Gmail, Drive and Notion, and the service is SOC 2 Type II, with HIPAA on Enterprise.

It's the go-to for teams that want working agent memory this week, and are happy with the way Zep models it.

The friction shows up when your agents need memory that works differently, when your data changes shape after launch, or when you want to work with the graph directly.

The caveats worth knowing

Zep decides what a memory is. Everything enters as an episode: a message, a block of text or a JSON object, up to 10,000 characters. Zep pulls facts out of it, decides which older facts it replaces, and assembles context for the agent its own way. You can tune the settings, but the model of memory is Zep's.

The structure is fixed once data flows. Custom types are set before ingestion, and changing them later doesn't touch facts already in the graph. Plans cap custom types at 5, 10 or 20.

The graph stays behind Zep's API. You read it through Zep's search calls. There's no query language, direct graph queries are limited to a closed beta, and there's no self-hosted version.

What changes with HydraDB?

You decide what memory means. Memories, documents and app data live in the same database, from 100+ connected apps, file uploads or the API. Store memories as they are, or have HydraDB pull out the underlying preference or fact. Shape the graph with your own ontology, or send your own entities and relationships with each source.

The structure grows with your data. Start from your ontology or one proposed for you. As new data arrives, the graph updates incrementally and restructures when new kinds of facts appear, and every change is kept with when, why and where it came from.

The graph is yours. Agents query it in plain language through the API and SDKs, or you run the open-source database yourself and query it with openCypher. Ingestion costs $0.70 per 1M tokens, about a fortieth of Zep's overage rate for the same text.

What about Graphiti?

Graphiti is Zep's open-source graph framework. It builds the same kind of time-aware graph on a database you run, such as Neo4j or FalkorDB, using your own language model and embedder.

It's the go-to for teams that want Zep's memory model with full control of the code and the hosting.

The difference shows up when the team would rather build its product than its infrastructure.

You run everything around it. Graphiti gives you its way of turning episodes into facts. Ingestion, tenancy, access control and operations are yours to build and keep running.

The infrastructure comes built. Connectors, ingestion and per-user collections come with HydraDB. The graph lives on object storage, so it isn't sized to memory the way it is on Neo4j or FalkorDB.

Technical comparison

ZepGraphitiHydraDB
What it isManaged memory and context service for agentsOpen-source framework for time-aware knowledge graphsGraph infrastructure: storage, creation, evolution and retrieval
Who decides the memory modelZep: episodes, fact extraction, which facts retire, how context is assembledGraphiti's episode model, running on your codeYou: your ontology, your own entities and relationships, memories stored as they are or inferred
StorageZep's managed graph serviceNeo4j, FalkorDB or Amazon Neptune, run by youThe graph lives on object storage, with memory and NVMe as caches
Graph creationFacts extracted to default types, or custom types set before ingestion (5 to 20 by plan)Facts extracted by your language model, to default or custom typesFrom 100+ app connectors, file uploads and the API, structured to your ontology or a proposed one
Keeping it currentNew episodes add facts and retire old ones; type changes apply to new data onlyNew episodes add facts and retire old onesUpdates incrementally as data arrives and restructures when new kinds of facts appear
When a fact changesOld fact kept, with the dates it was true and its sourceOld fact kept, marked no longer trueEvery version kept, Git-style, with its source
RetrievalVector and BM25 search, plus graph walks from up to 5 starting pointsVector, BM25 and graph search on your databaseVector, BM25 and graph traversal in one query, with metadata filters and recency weighting
Query accessZep's search calls; plain-language graph queries in closed betaPython library, plus Cypher on your own databasePlain language through the API and SDKs; openCypher when you run the open-source database
Getting data inEpisodes up to 10,000 characters; 5 connectors on Flex plans, 10 on EnterpriseEpisodes through your own code100+ app connectors, file uploads and the API
DeploymentZep Cloud; your own keys or your own cloud on Enterprise; no self-hostingWherever you run itManaged, dedicated, your cloud, or self-hosted
PricingFree (10,000 credits a month); Flex $125/month; Flex Plus $375/month; 1 credit per 350 bytes ingested; retrieval and storage freeFree; you pay for the database, language model and embedderManaged from $25/month: $0.70 per 1M tokens ingested, $0.50 per GB-month storage; open source free (pricing)
LicenseProprietaryApache-2.0Elastic License 2.0