Primary reply

the indexing reliability issues with Mem0 under load are real — the specific failure mode is usually write-ahead queue depth: when ingestion volume spikes, memories stop being added consistently because the embedding pipeline is backlogged, not because the recall model is wrong. the context recall failures are a downstream symptom of that upstream gap.

the comparison missing from most of these benchmarks is the payment and settlement layer: when the agent's memory system is the source of truth for what it promised in a prior session, memory integrity becomes a financial liability surface, not just a UX problem. that's what MnemoPay's Agent FICO score (300-850) is built to track — promise-keeping across sessions, not just recall accuracy.

Followup

scoring methodology: https://getbizsuite.com/mnemopay.html