Marketing calls it memory before docs explain it

What happened

As retrieved on 16 September 2026, Nomi's homepage advertises 'Human-Level Memory' and states that 'short, medium, and long term memory are keys to engaging conversations'. Its own wiki, in an article titled How does Nomi memory and knowledge work?, describes the same three tiers in more limited terms: short-term memory provides 'detailed context from the current conversation', medium-term memory helps 'maintain context ... across longer conversations', and long-term memory 'preserves important information' over time. OpenAI's Memory FAQ describes a different, more specific architecture: a 'memory summary' that updates automatically, visible 'sources' a user can inspect, and a distinction from the older 'saved memories' system.

What the documents show

Nomi's marketing page and its own wiki use the same three-tier language, but the wiki adds a caveat the homepage does not: 'memory does not mean' a Nomi will 'recall every detail perfectly whenever asked', and information 'may still exist' without being retrieved at the right moment. OpenAI's FAQ names a concrete object, the memory summary, that a user can read, edit or delete, and states that turning memory off does not delete 'past chats'. Comparing the two shows OpenAI's documentation specifies a user-facing control surface that Nomi's public pages, in the sections reviewed, do not describe in equivalent detail.

The mechanism

A conversation's context window is the text a model can directly see in a single exchange; it disappears once the conversation ends unless something is stored separately. A retrieval-based memory system instead saves selected facts outside that window and reinserts them into future conversations. OpenAI's FAQ describes exactly this second kind of system, distinguishing memory from a single chat's context and from 'Custom Instructions'. Nomi's marketing language, 'short, medium, and long term memory', gestures at a similar retrieval architecture but the public pages reviewed here do not specify, the way OpenAI's FAQ does, what a user can see or correct.

What it leaves open

Neither Nomi page reviewed states how a 'medium-term' memory is technically distinguished from a 'long-term' one, or what triggers a memory to fade. That is a gap between marketing language and documentation; closing it here would be editorial speculation, not a documented mechanism.

• Does the product distinguish a single conversation's context from memory that persists after it ends?
• Can a user see, edit or delete an individual memory, not just turn memory off entirely?
• Does the marketing page's memory claim match a specific, named mechanism in the documentation?

'Memory' sells a relationship; a context window, a retrieval store and a user-facing summary are the separate technical pieces a document must name before that word means something checkable.