Honest breakdown of AI companion app memory systems after testing 40 platforms
Memory is the most marketed and least delivered feature in the AI companion space. Most platforms claim to remember you but either reset between sessions or just pull from a profile you filled in manually. After two years of testing the ones that actually carry real conversational context across weeks are rare. Just published a full breakdown of which platforms actually deliver on this versus which ones are just marketing: medium.com/@companaya/nomi-ai-review-2026-is-it-worth-it-tested-c91811dcb24a
Any analytics on token usage?
Cool, but I can’t imagine how compute heavy it is to keep a running log of interactions and constantly include it in the context window and/or RAG. Week to week, sure, but over months to a year is wild, especially if you’re talking to it all day every day.
No public analytics that I've seen platforms keep that close to the chest. From actually using these things though the ones that hold up seem to be doing selective memory extraction rather than dumping the full conversation history into every context window. Key facts, emotional markers, relationship milestones — that kind of thing. Which makes sense because a raw running log would be insane compute costs for a $15/month product. Still way more expensive than session-only chat which is probably why free tiers get quietly capped so fast.
For sure. I can imagine these types of tools will keep getting bigger and leveraging massive amounts of data for a single user. There’s a future where the models are being tuned or even pre-trained (is that right?) on a regular basis. You’re in school. You get an AI assistant on your new laptop. It asks you some setup questions and tunes on the spot. After a month it pre-trains with your data and then continues to tune.
I was thinking it must be unrealistic and kind of freaky as an AI companion to remember so many little things. As an assistant though, it’s great.
Llama 3-based stuff is 70B, but it's not the largest out there. Not even the largest open-weight model. Off the cuff, Behemoth is based on Mistral Large, at 123B.
its so weird that that is now mistrals "medium" size...
weird to see em growing up like this...
Right? A year ago 70B felt massive and now it's basically mid-tier. The pace of this is genuinely hard to keep up with even when you're paying close attention to it.
And over here, I thought was 380 billion. Parameter model was fucking small.
Fair point and you're right I should have said largest commonly deployed in consumer companion apps rather than largest period. Behemoth at 123B exists but none of the platforms I tested are actually running that in production. Sloppy phrasing on my part, appreciate the correction.