Theory-of-Mind forgetting curves

What does the user still remember? Each item has a stability that grows logarithmically with exposure count and linearly with reinforcement, then decays as R(t) = exp(-t / S). The plot extends each curve from now to the horizon. Use it to decide when to re-surface a fact before it falls below useful recall.

Formula and parameters ported from src/relevance/tom/forgetting-curve.ts in the cortex-ai repo.

Negative-space retrieval

Given the entities currently active in your context, surfaces other entities that historically co-occur with them but are absent right now. Each finding's significance weights raw co-occurrence by recency, so a 30-day-stale partner is suppressed but a fresh absence is loud.

Algorithm ported from src/relevance/negative-space/absence-detector.ts. The co-occurrence table and last-seen ledger here are curated stand-ins for what the live cortex-ai SQLite database would hold for your machine.

Hippocampal consolidation

Compresses a stream of fine-grained episodes into a smaller set of semantic capsules. Each capsule keeps the TF-IDF spine (tags + verbatim spine sentences) of the episodes it groups, so downstream agents read a few capsules instead of dozens of episodes.

TF-IDF distillation ported from src/relevance/consolidation/distiller.ts. Clustering uses scikit-learn agglomerative clustering on cosine distance.