Who builds your graph: an LLM pipeline or an on-device model?
Cognee is graph-first: it turns documents, code and conversations into a knowledge graph you can shape with ontologies. Dakera builds its entity graph as you store memories, with a model that runs inside the server. The deciding question is what does the extraction work, and what you want to own.
Reviewed October 2026
From document to graph
Cognee's pipeline has three stages: add or remember, cognify or improve, then search or recall. With an LLM key configured, Cognee uses OpenAI by default and processing makes provider calls. Without one, its keyless mode builds the graph with a local GLiNER model and a local embedding model, and LLM-dependent improvement stages skip.
The practical difference is in the default path. Dakera's extraction is always on-device, enabled per namespace, so the cost of ingesting more documents is your own server's compute and nothing per call. In Cognee's LLM mode, graph quality follows the model you choose, which is its strength for rich, domain-specific relations.
Ingest beyond text
Cognee accepts text, code, conversations, database records, audio and images, and recent releases add Google Drive and Gmail connectors. For audio and images, its README says a vision or transcription model needs LLM configuration. Dakera v0.12.0 handles media with models that ship with the server. All of it is opt-in.
| Input | Cognee | Dakera v0.12.0 |
|---|---|---|
| Text and conversations | Yes, with or without an LLM key | Yes; sessions, agents and namespaces built in |
| Audio | Supported; needs an LLM configured (per the Cognee README) | On-device speech to text; five Whisper models, multilingual whisper-base by default, language auto-detected |
| Document pages and images | Supported; needs an LLM configured (per the Cognee README) | Visual recall over document pages from the text of a question, on-device |
| Languages | Not published | bge-m3 covers 100+ languages in one space; full-text stemming in 18 plus CJK bigrams; dates in 7 |
| Connectors | Google Drive and Gmail in v1.6.1 | Attachments you upload |
Ontologies, and what Dakera keeps in the graph
If your domain demands custom entity types and relations, Cognee's ontology and data-model support is the real draw, and Dakera does not try to match it. Dakera's graph is deliberately simple: entities extracted by GLiNER, five edge types, breadth-first traversal, and edges private to each agent by default. It sits beside hybrid retrieval, memory decay, TTL and spaced-repetition extension, so recall can use the graph without making the graph the whole product.
Running it
- Cognee: a Python package or the
cognee/cogneeDocker image, plus Cognee Cloud as a managed option; Apache 2.0. - Dakera: one container, 783 MB for the CPU image with default models baked in and ready about 4.4 s after start, so it also starts air-gapped. MIT SDKs and MCP server; proprietary server binary with no usage fees.
- Operations in v0.12.0:
/health/ready,dakera --check-config, encrypted backups, AES-256-GCM at rest once a key is configured, multi-node high availability with replicated writes, and one-command rollback withdakera downgrade.
When to choose each
Choose Cognee if
- Your product is a knowledge graph, and you want custom ontologies and data models.
- LLM-quality relation extraction over your own corpus is worth the provider calls.
- You already operate Neo4j, Neptune or Memgraph and want memory in that graph.
- You want Google Drive or Gmail ingestion in the SDK, or a Python library inside your app.
Choose Dakera if
- You want a memory server with sessions, decay, TTL and hybrid recall, with a graph included.
- Extraction must run locally with no per-call cost.
- Your data includes speech, document pages or several languages, and must stay on your hardware.
- You need SDKs in Python, TypeScript, Go and Rust and an MCP server for Claude Desktop, Cursor and Windsurf.
Prefer managed? Dakera Cloud is coming. Join the waitlist.
Frequently asked questions
Does Cognee need an LLM?
Not for text. Cognee's README says text ingestion, retrieval and session storage work without an LLM using a local GLiNER model and local embeddings, while generated answers and audio or image processing need LLM configuration.
Does Dakera use an LLM to build its graph?
No. Entity extraction runs on-device with GLiNER, and embeddings, reranking and speech to text run on-device as well. No LLM or external API sits in the memory path.
Can Dakera model custom ontologies?
Not in the way Cognee does. Dakera extracts entities into five edge types, and you can filter by entity types per namespace. Choose Cognee when a custom ontology is central.
Which has the better graph?
They aim at different things. Cognee optimises for rich, configurable knowledge graphs, and Dakera treats the graph as one signal beside hybrid retrieval, decay and sessions.
Run Dakera on your own infrastructure.
One binary with its models baked in and no per-call fees. Start it with one Docker command and see it for yourself.
Prefer managed hosting? Join the Dakera Cloud waitlist →