Memory

Brand memory keeps what must survive between conversations. It follows the model the large labs have converged on: a small core always present and the rest retrieved on demand, with people deciding what is permanent.

How something becomes memory

1. Nödo or the background extractor proposes a fact or a decision, citing the message, task or file.
2. Extraction only reads shared conversations; never internal or private threads.
3. A person accepts, corrects or dismisses it in Genome › Memory.
4. On acceptance it is checked against accepted memory: what it replaces moves to history with an end date, never deleted.
5. Accepted memory is indexed for keyword and semantic search.

Core and episodes

• The core is compiled without a model, always the same for the same inputs, and recompiled whenever accepted memory, the Genome or an episode changes. Nödo reads it at the start of every conversation, call and summary.
• Whatever does not fit in the core stays in memory, and Nödo retrieves it with memory.search when needed.
• An episode summarises a shared conversation once it has been quiet for two hours or is closed, and each work once it is completed. It is never made from internal or private threads, and when a conversation stops being shared or is archived, its episode goes away.
• The brand sees the core, when it was prepared and its episodes, each linked to its conversation or work, in Genome.

Search

memory.search combines keyword and semantic search, fuses the results and returns each hit with its source. It accepts asOf to ask what was true on a given date.