Nomi’s March 2025 memory update emphasized continuity over a larger context window

The release claim

Nomi’s March 11, 2025 update said its memory system had expanded capacity and improved medium-term retention. It specifically claimed clearer recall of events more than 1,000 messages earlier, better handling of story elements, and fewer hallucinated, conflated, or misunderstood details. These are vendor claims about a deployed system; they are not a benchmark, guarantee, or independent reliability result.

The prior Identity Core announcement places that release in a broader architecture. Nomi described short-, medium-, and long-term memory, backstory, shared notes, and an identity mechanism that can retain material it considers important to the character. That design separates a factual recall function from a representation of the companion’s own continuing persona. The distinction matters in roleplay or long-running conversations, where a system may need both a remembered event and a stable way of responding to it.

How to read the claim

Evidence: the March post dates and names the claimed improvements; the December post describes the surrounding memory vocabulary. Inference: ‘continuity’ is the user-facing goal, but it can fail in several ways—missing a fact, retrieving a stale fact, or applying a remembered fact in the wrong context. A larger retained history does not itself solve those selection and interpretation problems. The announcements do not provide error rates, a data-retention policy, or a user-auditable log of all recall decisions.

Evidence limit: No independent testing was conducted; performance language is attributed to Nomi.