Cognee vs Supermemory: Which AI Memory Tool to Run
On this page, 12 sections
Cognee and Supermemory both claim to give an AI agent a memory, both are open source, and both sit at about 31 to 32 thousand GitHub stars. That is where the similarity stops being useful.
One builds a knowledge graph from your data. The other sells a memory API with a local binary on the side. This post compares what their own READMEs, docs and pricing pages say, read on 11 October 2026.
I have not run either one in production, so I am not going to tell you which one “felt” better. Every number below comes from a page I opened that day, and where a page says nothing I say so.
The short answer
Pick Cognee if you want a graph you can inspect and you are fine running Python. Pick Supermemory if you want a hosted memory API with an MCP server and a monthly plan with included credits, and you would rather not manage a graph database.
Both document a local setup. Supermemory’s README and self-hosting docs describe a fully offline mode with Ollama and local embeddings, so the choice there comes down to graph versus memory API.
What each one actually is
Cognee is a Python library. Its docs describe a knowledge graph that you fill with remember and query with recall. The GitHub repo is topoteretes/cognee, licensed Apache-2.0, with 32,009 stars when I checked the GitHub API. The last push was 9 October 2026.
Supermemory is a memory and context service. The repo is supermemoryai/supermemory, licensed MIT, with 31,185 stars. The last push was 11 October 2026, the day I read it. It ships an MCP server, and the README has a section called “Supermemory local” for running it yourself.
For the wider field, including Mem0, Zep and Letta, I keep a separate overview in AI memory tools compared.
Side by side
| Item | Cognee | Supermemory |
|---|---|---|
| License | Apache-2.0 | MIT |
| GitHub stars (11 Oct 2026) | 32,009 | 31,185 |
| Last push | 9 Oct 2026 | 11 Oct 2026 |
| Core idea | Knowledge graph, remember and recall | Memory and context API |
| Free tier | $0, 1 workspace, 1M tokens | $0, $5 of credits a month |
| Paid entry | $1.00 per 1M tokens, plus $5 per extra workspace | $19 a month, $20 included |
| Run it yourself | Yes, Python library and a Docker image | Yes, “Supermemory local” |
| MCP | Plugins and MCP listed in the README | Yes, MCP server |
Pricing, read on 11 October 2026
Cognee’s pricing page lists three plans. Free is $0 with one workspace and 1M tokens. Standard is $1.00 per 1M tokens plus $5 for each additional workspace. Enterprise is a bring-your-own-cloud engagement, so there is no public number.
Supermemory prices in credits. Free is $0 with $5 of credits each month. Pro is $19 a month and includes $20 of credits. Two bigger tiers sit above it: $100 a month with $130 included, and $399 a month with $600 included. Credits are spent in SM tokens, which cost $5 per 1M for plain text and $10 per 1M for rich content.
Read those two together. Cognee charges per token plus per workspace, so your bill follows how much you ingest and how many separate memories you keep. Supermemory charges a monthly fee that includes credits, so a small team knows its bill in advance. I cannot tell you which is cheaper for you without your token volume. Work that number out first.
Running it without paying anyone
This is the part where the two differ most in what the docs say.
Cognee’s README has a section titled “Run locally without an LLM”. You add the package with the gliner extra. With no LLM key set, Cognee extracts the graph with a local GLiNER model and embeds with a local embedding model.
Both download on first use. The README says text ingestion, retrieval and session storage work without an LLM, and that LLM-dependent improvement stages skip automatically.
There are two limits in that same README. Generated answers need an LLM configured. And the bundled GLiNER extractor is described as a demo of the small-model pipeline, with a production-grade version available by contacting the company. So “free and local” gets you a working graph and retrieval, not the full pipeline.
Supermemory’s README gives a one-line setup, a curl script or npx supermemory local, and then you run supermemory-server. The docs have a self-hosting overview page.
The README and the self-hosting docs say it runs fully offline, using Ollama and local embeddings, so no outside API key is needed for that mode. I have not run it myself, so treat the offline claim as the project’s own, read on 11 October 2026.
The benchmark claims
Supermemory’s README says it ranks first on LongMemEval, LoCoMo and ConvoMem, and claims 95% Recall@15 with a 99.4% context reduction. Those are the vendor’s own claims, and I have not reproduced them. Treat them as a reason to run the benchmark on your own data, not as a result.
Cognee’s pages I read do not put a head-to-head benchmark number up front. That is not a weakness or a strength by itself. It means there is no vendor number to compare, so you would be testing both on your own questions.
Where I would pick Cognee
If your data is documents, tickets or notes and you want to see how entities connect, a graph is the right shape. You can look at the nodes. When retrieval is wrong you can see why. It is also the more direct fit if you want to start free on your laptop with the open source library.
If you are comparing Cognee against other graph-first tools, my notes are in Cognee alternatives.
Where I would pick Supermemory
If you want to wire memory into an assistant through MCP this afternoon and pay a known monthly price, Supermemory’s shape is easier. The pricing is simple to explain to a client. The MIT license is shorter than Apache-2.0, though for most readers both allow commercial use and modification. Check each license text before you build a product on it.
If you are still deciding which kind of memory layer you need at all, start with the best AI memory layer for agents.
What I could not verify
- How well Supermemory’s local mode performs offline. The README says it runs fully offline with Ollama, and I did not run it.
- Real-world cost for either tool at your volume. Both prices depend on usage.
- The benchmark rankings. They are self-reported.
- Anything about uptime or support quality. I did not test either service.
Do this today
- Write down 20 real questions your agent should answer from memory.
- Run Cognee locally with the
glinerextra and ingest the same documents you would give Supermemory. - Run
npx supermemory localand ingest the same documents. - Ask both the 20 questions. Count correct answers, not impressions.
- Estimate your monthly tokens, then price them against both pricing pages.
Frequently asked questions
Is Cognee free?
The open source library is Apache-2.0 and the pricing page lists a free plan at $0 with one workspace and 1M tokens. Running it locally without an LLM key is documented in the README.
Is Supermemory open source?
Yes. The repo supermemoryai/supermemory is MIT licensed, and the README describes a self-hosted “Supermemory local” option.
Which one has more GitHub stars?
Cognee, with 32,009 against 31,185 for Supermemory on 11 October 2026. The gap is under 3 percent, so it tells you little about quality.
Can I use either with MCP?
Supermemory has an MCP server. Cognee’s README lists plugins and MCP under connecting an agent.