The honest comparison
| Question | Zep | MongrelDB |
|---|---|---|
| What is it? | Zep Cloud is a managed agent-memory platform. The public getzep/zep repo is examples and integrations, not the product. | An open source embedded database engine written in Rust. |
| Open-source path | Graphiti, a temporal knowledge graph framework, is the related open-source project. | MongrelDB itself is MIT or Apache-2.0 licensed. |
| Storage dependency | Zep Cloud uses its vendor-run context engine. Graphiti requires an external graph database such as Neo4j, FalkorDB, or Neptune. | No required external database. Run embedded, through native bindings, or as a local daemon. |
| Model dependency | Graphiti requires an LLM API key for extraction and embeddings. Zep Cloud manages the pipeline as a service. | Core storage is embedding-vendor neutral. Applications supply vectors, register providers, or use sparse retrieval without a dense model. |
| Retrieval model | Temporal knowledge graph with entity, relationship, validity-window, semantic, and keyword retrieval. | Dense ANN, sparse retrieval, FM substring, bitmap equality, learned range, and MinHash indexes intersecting in one RowId space. |
| Best fit | Teams that want managed temporal memory assembly. | Teams building their own memory layer on owned storage. |
Where Zep is strong
Zep's public materials position it around temporal knowledge graphs: entities, relationships, episodes, fact validity windows, invalidation when new data supersedes old facts, and prompt-ready context blocks. That model is directly aimed at agent memory, and the managed service removes the storage and pipeline work.
Graphiti is the important open-source counterpart. It is Apache-2.0 and gives developers a Python framework for building temporal knowledge graphs, but it is not a self-contained embedded database. Its README lists external graph database requirements and an LLM API key. If you want Zep's worldview without Zep Cloud, Graphiti plus that infrastructure is the path to evaluate.
Where MongrelDB is different
MongrelDB starts lower in the stack. It gives you the durable memory store and the retrieval machinery, while your application owns the memory policy. That is a better fit when the agent's memory is product state, not a detached service call.
MongrelDB can represent temporal behavior without adopting a graph database: store timestamps, valid-from and valid-to fields, current state, reinforcement counts, importance, project, entity, and memory type as indexed columns. Learned-range and bitmap indexes enforce the temporal and metadata boundary; dense, sparse, FM, and MinHash indexes handle recall and duplication.
The Hermes memory plugin is a concrete example. It uses MongrelDB-backed memory in native FFI or daemon mode, defaults to dense ANN with all-MiniLM-L6-v2 at 384 dimensions, offers model-free sparse retrieval, and creates encrypted data directories by default.
Choose based on operating model
Choose Zep or Graphiti when
- you want temporal knowledge-graph memory as the central abstraction;
- a managed memory service is acceptable or preferred;
- you are willing to run Graphiti's graph database and LLM dependencies for the open-source path;
- your team wants memory assembly handled by the framework.
Choose MongrelDB when
- agent memory must stay local, embedded, or inside your service;
- you want SQL and transactional writes around memory rows;
- retrieval must mix semantic, exact, metadata, recency, and dedup signals natively;
- encryption at rest and credentials are required;
- you do not want every memory write and search to depend on an external LLM or third-party memory service.
Evaluation checklist
- Model a memory correction: old fact invalidated, new fact current. Test whether the store can retrieve the current state without graph ceremony.
- Query by time range, project, user, exact phrase, and semantic paraphrase in one request.
- Count external dependencies on the write path and search path. Each LLM, graph database, or hosted service is another availability and privacy boundary.
- Test deletion and retention. Agent memory needs a forget path as much as a recall path.
- Compare operational fit, not only answer quality: backups, encryption, credentials, local development, offline tests, and disaster recovery.
Sources
Zep comparison FAQ
Is Zep still self-hosted?
Zep's current product is Zep Cloud. The old Community Edition is deprecated, and the getzep/zep repository now carries examples and integrations. The open-source framework to evaluate is Graphiti.
What does Zep do better than MongrelDB?
Zep gives you a managed temporal memory platform. Graphiti gives you a framework when you want its temporal knowledge-graph model and accept the external graph database and LLM requirements.
What does MongrelDB do better?
MongrelDB owns the storage and retrieval layer locally: hybrid indexes, SQL, transactions, encryption, and daemon or embedded deployment without a required third-party database.
Is Graphiti the same thing as Zep?
No. Graphiti is the open-source temporal knowledge-graph framework. Zep Cloud is the managed service built around Zep's context engine.
Can MongrelDB represent temporal memory?
Yes, through schema design. Index timestamps, validity windows, state, project, entity, and importance, then combine those filters with dense, sparse, exact, and dedup retrieval.