Part 1 of 6 · 1 min read
Overview
Module 2 decided what to store. Module 3 decided what to do when the history grows. This module is the read path of long-term memory: the store may hold thousands of items, the window holds a handful, and the agent needs the right handful.
Level 4 covered retrieval over documents (RAG). Memory retrieval differs in four ways that this module is about: memories change (a fact has a current and a past value), they carry time and importance, they have an author whose trust matters, and a wrong memory persists into every later session.
About the example, and what is real
Everything runs offline in code/mini-recall. Real: the store, the write path (gate, supersession, hard delete), the scoring formula, the four policies, the judge, the 69 tests. Imitated: the embedding. A real embedding model is a learned dense vector; here it is a bag of concepts built from a small hand-written synonym table, weighted by inverse document frequency, compared by cosine similarity. It reproduces the behaviour that matters here (paraphrases match, near-duplicate values score the same), but its numbers are not those of a real model. The world is generated: 5 keyed facts (3 revised later), one important constraint, 100 chatty notes, 6 questions per world, 5 worlds. The score weights are my choices; I did not tune them to flatter the result, and the weight sweep shows what happens when they move. No real model is called.