Disclosure: I've been exploring CortexDB (cortexdb.ai) for a project, so the write-path question is top of mind. Bayesian belief updating is an interesting angle for confidence-weighted retrieval, but the question I'd push on is: when a belief updates, does the underlying memory get rewritten, or does the system append a new observation and recompute the posterior? CortexDB takes the latter approach — append-only events, beliefs derived at query time — which preserves the raw signal so you can re-derive beliefs whenever your model improves. Typed atoms (episodic/procedural/semantic) face the same issue: once you've committed to a classification at ingest, walking it back is hard. How does Mnemo handle re-derivation when the belief model improves?
I work with small clinics in India — their entire patient "CRM" is WhatsApp scroll-back. Doctor replies "ok do MRI" with no context, patient's son follows up two days later from a different number asking about cost, receptionist forwards a screenshot of the fee chart from another group. That's what one patient thread looks like, and there are hundreds.
Going to try piping a WhatsApp chat export from one of these clinics into CortexDB this weekend. The stuff I want to query is mundane — what was this patient's last X-ray finding, did we already discuss surgery with the family, how many ACL cases did we see this quarter. Stuff the receptionist currently answers from memory or by scrolling for 10 minutes.
The event-sourced approach is what makes this interesting to me over something like Mem0. This data is messy enough that I don't trust any model to extract the right structure at write time. Store the mess, derive views later, re-derive when the model gets better — that feels right for this kind of input.
Will report back.