Kognite vs Mem0: where the graph paywall sits.

Mem0 bills adds and retrievals as separate meters — the free tier is 10,000 adds but only 1,000 retrievals a month. Retrieval is the operation an agent does most (every proactive recall is a retrieval), so teams burn the retrieval budget first, often without noticing which counter emptied.

Kognite counts one thing: memories created. Searches and retrievals are not metered on any plan. The practical consequence: you can tell your agent to recall proactively on every prompt — the pattern that makes memory actually useful — without the recall itself consuming quota.

We ran Mem0 (OSS 2.0.11) through our own public benchmark harness on the LoCoMo corpus and published both sides' numbers — retrieval recall and end-to-end answer accuracy, same questions, same judge. The run, the harness, and the caveats are on the benchmarks page; we'd rather show the measurement than adjectives.

• ✓You want to self-host. Mem0's open-source library runs in your own infrastructure with your own vector store. Kognite is a managed service only.
• ✓You're building on their SDK ecosystem. Mem0 has a large community and many framework integrations; if your stack already embeds it, switching costs are real.
• ✓You need graph memory and have the Pro budget. At $249/mo, Mem0's graph with entity linking is a capable product — the argument here is the price of admission, not that it doesn't work.