Novelty
A retrieval weighting signal computed as 1 - decay^(t - lastSeen(id)). Memories not recently accessed during the current session score as novel, promoting retrieval diversity and preventing attentional fixation on the same high-importance memories.
Novelty in the retrieval context is session-relative: a memory is novel if it has not been accessed recently during the current session, regardless of how old it is or how often it has been retrieved in past sessions. The formula 1 - decay^(t - lastSeen(id)) produces a score near 1 for memories not seen in this session and a score near 0 for memories retrieved moments ago. This prevents a common failure mode in attention systems: fixation on a small set of high-importance memories that score well on every retrieval and crowd out others. By giving recently-retrieved memories a low novelty score, the retrieval pipeline is pushed to surface a broader range of relevant evidence across a session. The novelty signal is also modulated by usage frequency: the combined novelty score used in reinforcement is 70% raw novelty plus 30% inverse usage frequency, so a memory that is retrieved very frequently across sessions does not receive a full novelty bonus even when it has not been accessed in the current session. This prevents high-frequency memories from continuously collecting novelty credit in addition to their usage-frequency credit.