Mnemonic Threads: A Dual‑Memory Architecture for LLM Continuity

This document proposes Mnemonic Threads, a dual‑memory architecture for sustaining continuity in large language model (LLM) interactions. The approach separates short‑term working context from an append‑only mnemonic ledger that stores salient anchors, citations, and thread markers. Retrieval is guided by thread‑level signals and lightweight heuristics to minimize token usage while preserving provenance and reducing hallucination. Prototype notes, experiment outlines, and deployment artifacts are included to demonstrate feasibility and to provide runnable specs for reindexing, thread tracking, and provenance thresholding. The architecture targets interactive assistant scenarios where long multi‑turn continuity, low latency, and strong attribution are required.