MindForm — Memory: Two Retrieval Systems, One Never-Forgetting Design

Every substrate documented so far in this series runs as a stage in the turn: a refresh, an event, a formation step. Memory does not, and the first finding of this paper is about its own subject rather than any single mechanism within it: Memory is not one node but two parallel systems — episodic, over a complete log of lived experience, and semantic, over the belief store documented in Character — each split into a human-readable record and a compact embedding sidecar, called into from four separate points in the turn rather than occupying a stage of its own. What counts as "the same kind of experience" is decided by a specific representational choice: cosine similarity over a MiniLM sentence embedding, the identical representation that grounds the architecture's own offline appraisal model elsewhere in the codebase, employed here for topical and phrasal similarity rather than felt similarity. Retrieval layers two further judgments strictly inside a hard eligibility floor — an emotional-vividness term and, where available, a retrieval-decay term that dulls old, unremarkable memories while protecting vivid ones almost entirely — with the displayed similarity always kept honest regardless of how the ranking was reweighted. One retrieval pass, computed once per turn, is read by three independently documented nodes for three unrelated purposes, with the consequence that three biases across the architecture can disagree about interpretation but never about evidence. Nothing is ever deleted from either system; this is a stated commitment rather than a missing feature, defended on the grounds that a memory's value lies in its detail and that the correct long-term answer to storage growth is compression, not forgetting — a position for which no implementation yet exists, alongside a retrieval mechanism that scales linearly against a character's complete history on every call. We identify four first-class quantities, across four consecutive nodes including this one, that the architecture computes with evident deliberateness and structurally cannot act on, and trace an expectation-confirmation loop whose reach is bounded by literal text similarity rather than by anything affective — a constraint we cannot determine to be protective or merely incidental.