Methodology
Why some competitors are not tested in max mode
The methodological rationale: LLM cost, ingestion time, local configuration, and the difference between architectural efficiency and a general ranking.
Max mode has a cost
The obvious criticism is valid: some systems were not tested with every intelligent option enabled. Mem0 runs with `infer=False`, Hindsight has no LLM provider, and Mnemosyne is FTS-only.
But that choice is neither free nor hidden. The June 10 report estimates that enabling LLM extraction on roughly 41 million tokens would cost around 50 to 100 dollars with a cheaper model, and up to around 445 dollars with Sonnet, with 8 to 25 hours of ingestion for some configurations.
What I am trying to measure
I am first trying to measure architectural efficiency: how much context the model needs to answer correctly, and how long it takes to ingest the corpus.
If a system needs expensive LLM extraction before it turns competitive, that cost is part of the product, and it does not get to hide outside the benchmark.
The limitation I acknowledge
Treat this as a series of tests on structured operational memory, not a universal ranking. It favors an architecture that can address entities and state directly, and I say so up front.
That is why each comparison page includes a caveat box, raw links and the rerun command.