Cognee
The memory engine for AI agents - knowledge graph + vector memory in one self-contained container.
One-click deploy, from $13/mo on a Miget plan.
Agents without memory re-learn your world every session. Cognee is the maintained open-source answer (Apache-2.0, 17k+ stars, releases this month): ingest documents and conversations, "cognify" them into a knowledge graph plus vector index, and query the result as long-term memory over a clean API.
The engineering taste shows in the defaults: SQLite, LanceDB, and Kuzu all run in-process - one container, one volume, no sidecar fleet. Every layer has a scale-out path via env (Postgres, the catalogue’s qdrant, Neo4j) when you genuinely need it.
It is API-first (Swagger at /docs; the upstream UI is a separate work-in-progress) and slots into the catalogue’s agent lane: route extraction through the litellm gateway, and give agent-box or OpenClaw agents a memory that survives them.
Upstream project: Cognee
#what you get
- Knowledge graph + vector memory from documents and chats
- Self-contained defaults: SQLite + LanceDB + Kuzu in-process
- Real auth (JWT users + API keys) - register, token, go
- OpenAI-compatible gateway support (litellm plugs in)
- Scale-out env paths: Postgres, Qdrant, Neo4j per layer
- Apache-2.0, very actively maintained
#topology
| Service | Role | Public |
|---|---|---|
| cognee | memory API + Swagger at /docs (:5000) | yes (JWT auth) |
#miget sizing
// this stack needs
2 GiB RAM · 10 GB disk · 1 service
Cognify jobs are the memory consumers (graph building in-process) - 2 GiB is the floor, heavy corpora want more. Upstream caps its own compose at 8 GiB.
Hobby - recommended fit
$13/mo
1 vCPU · 2 GiB · 50 GiB disk
Headroom for your own apps: 2 GiB at $19/mo
Professional - production
$22/mo
1 vCPU · 2 GiB · 10 GiB disk
Dedicated resources, production SLOs - plan details
One Miget plan is a fixed pool of compute - the whole stack (managed databases included) deploys inside it, and anything left over runs your other apps. No per-service or per-seat math.
#vs. other PaaS
Estimated monthly cost of running this exact stack (2 GiB RAM, 10 GB disk, 1 container) elsewhere, from published June 2026 rates.
| Platform | Est. monthly | Notes |
|---|---|---|
| Miget ★ | $13 flat | compose stacks first-class: one deploy, dedicated vCPU, managed Postgres/Valkey, volumes and TLS all included in the plan |
| Heroku | ~$100 | no volumes; nothing between 1 GB ($50) and 2.5 GB ($250) dynos - 2 GB containers cost far more than shown |
| DO App Platform | ~$29 | no persistent volumes - stateful containers need managed DBs/Spaces (base $5 Spaces included here) |
| Render | ~$28 | per-service instances (0.5 GB $7, 2 GB $25) - every container is its own paid service |
| Railway | ~$22 | usage-based ($10/GB RAM-mo); vCPU billed separately at $20/vCPU-mo on top |
| Fly.io | ~$13 | cheapest sticker price - but burstable shared CPUs (1/16 core; dedicated vCPUs cost ~2-3×), no compose deploys (one app per container, manual wiring), managed DBs billed extra |
Estimates assume RAM fully allocated at published on-demand rates - and sticker price isn't the whole comparison: the cheaper rows buy burstable shared CPUs, per-service wiring instead of a compose deploy, and managed databases billed separately. Heroku and DO App Platform have no persistent volumes at all - stateful stacks like this one need workarounds there.
#deploy it
On Miget
- Create a Compose Stack in app.miget.com pointing at the templates repository
- Set the stack path to
cognee -
Set the required variables:
LLM_API_KEY, fact extraction + embeddings (billed to you; litellm virtual keys work)FASTAPI_USERS_JWT_SECRET, overrides an insecure default
- Deploy. Miget layers
compose.miget.yaml(RAM, privacy, volumes, managed services) automatically
Locally first?
Every template is portable, vanilla Docker Compose - the Miget overrides are ignored locally:
git clone https://github.com/deployable-sh/stacks
cd miget-compose-templates/cognee
docker compose up -dSame files, same behavior. The template README covers connection strings and scaling notes.
#faq
Cognee or Mem0 for agent memory?
Mem0 is the bigger name, but its published self-host images are stale (the current server has no amd64 image - we verified before choosing). Cognee ships current images, runs self-contained, and adds the knowledge-graph layer plain vector memory lacks. When Mem0 publishes proper images, it earns a template too.
How do agents use it?
Over the REST API: add content, trigger cognify, then search with graph-aware retrieval - returning facts and relationships, not just similar chunks. Register a user, mint a token, and any agent in the project calls http://cognee:5000.
What do the LLM calls cost?
Cognify extracts entities and relations via your LLM and embeds via your embedding model - that is the real cost, scaling with ingested volume. Route through the litellm template to cap it with a budgeted virtual key, and consider the tei template for the embedding leg.
Ship Cognee today
One compose stack, 2 GiB of RAM, from $13/month flat, and it runs on your laptop with the same files.