Memory as Metabolism: A Design for Companion Knowledge Systems

A new class of personal LLM memory systems emerged in April 2026 — Karpathy's LLM Wiki, MemPalace, and LLM Wiki v2 — which compile knowledge into a persistent interlinked artifact rather than retrieving from raw documents on every query. These systems are a genuine step forward. They share a gap: none specifies what a companion memory system is obligated to do, what it must not do, and what failure looks like for the specific problem of entrenchment under user-coupled drift. This paper proposes a governance profile for that gap. It sits alongside 2026 governance frameworks like Context Cartography and MemOS — not claiming to replace them, but specializing their logic for single-user companion wikis where the success metric is stable user utility under drift, not objective truth-tracking. The core design principle: personal LLM memory is a companion system. Its job is to mirror the user on operational dimensions (working vocabulary, load-bearing structure, continuity of context) and compensate on epistemic failure modes (entrenchment, suppression of contradicting evidence, Kuhnian ossification). Five operations implement this split: TRIAGE, DECAY, CONTEXTUALIZE, CONSOLIDATE, AUDIT — supported by memory gravity and minority-hypothesis retention. The paper includes an object model with defined lifecycle states and status flags, normative conformance invariants specifying must/must-not obligations for each operation, and worked traces demonstrating legal and illegal system behavior. The sharpest prediction: accumulated contradictory evidence should have a structural path to updating a centrality-protected dominant interpretation through multi-cycle buffer pressure accumulation — a failure mode no existing benchmark captures. This version (v3.64) adds two structural pieces to the earlier draft: a conflict routing matrix that specifies how the mirror-vs-compensate principle routes when operational continuity and epistemic correction point opposite ways, and a formal treatment of memory gravity with a structural base component and an access-modulated effective component, including the four properties a conforming gravity function must satisfy. An application context for AI-assisted software development is added in §10.4. The mechanisms are mostly borrowed from cognitive architectures, belief revision theory, and graph centrality. The contribution is the companion-specific normative profile: what a single-user wiki must do when faced with user-coupled drift, assembled around a specific design target no prior framework explicitly addresses. The safety story at the single-agent level is partial, and the paper is explicit about what it does and does not solve. It is intended as a design reference that practitioners — Karpathy, MemPalace authors, or anyone building in this space — could draw on.