A knowledge graph vs a knowledge base
Zep is a context-engineering platform for agent memory, built on a temporal knowledge graph. Its open-source engine, Graphiti, turns conversations and data into a graph of entities and relationships that changes over time, so production agents can retrieve temporally-aware context. This is developer infrastructure aimed at scale, not a notes app.
Hjarni is a knowledge base you write in. Notes are plain Markdown, organized in folders, tagged and linked, and the built-in MCP server lets ChatGPT, Claude, and other clients search, read, and update them. The author is a person, and the unit is a note, not a graph node.
Who writes, and at what scale
With Zep, your agent writes into the graph as it runs, and Zep's value is the retrieval and the temporal reasoning over everything it has ingested. That is exactly what you want when an autonomous system needs production-grade memory.
With Hjarni, a person writes the note and any connected AI reads it. There is no graph to model and no ingestion pipeline. You write a runbook or a decision log, and the next time your assistant asks, it reads what you wrote.
Zep is graph memory for agents at scale. Hjarni is the notes a person actually writes.
Different buyer, different price floor
The price tells the story of who each product is for. Hosted Zep is a closed SaaS with a free tier (reportedly around 10,000 credits a month and 2 projects), then paid Flex from $125/mo and up to Enterprise with BYOC and VPC options. To self-host, you now run the open-source Graphiti framework plus a graph database yourself, since the old Community Edition is deprecated. Hjarni starts free at 25 notes, with Pro from €9 a month. One is priced for engineering teams putting agents in production; the other for a person who wants their notes to have a memory.
Verbatim notes vs extracted facts
The temporal knowledge graph is Zep's whole point, but it has a quieter cost. Because an LLM extraction step re-interprets each conversation into graph nodes and edges, the thing Zep stores is not what was said, it is the model's reading of what was said. That extraction can be wrong: it can record a confidently-timestamped "fact" that the user never actually stated, and once it is a node in the graph it reads back as if it were ground truth. The case is not that Zep is careless, it is that any re-interpretation step can drift from the source, and a graph hides that drift behind clean structure.
Hjarni stores the verbatim Markdown you or your AI wrote. There is no extraction layer between what was saved and what comes back, so you read exactly what was put there, and a named note can be opened, checked, and reverted. The tradeoff is real and worth saying plainly: you give up the temporal graph reasoning Zep does, and you get back memory you can trust to be your own words.
When Zep is the better fit
If you need enterprise-grade, temporally-aware knowledge-graph memory for production agents, and you have the engineering to run it or the budget for the hosted tier, Zep is built for exactly that. Graph RAG with temporal awareness is a real capability that a full-text notes app does not try to match.
When people choose Hjarni instead
The switch happens when the job is "I want to write notes my AI can read", not "I am running agents in production that need graph memory". Hjarni is the knowledge base and the MCP server in one hosted product, EU-based, with folder-level AI instructions and a free tier. No graph to model, no graph DB to operate. You write a note, your AI reads it.