Memory as Metabolism: A Design for Companion Knowledge Systems

Researcher Stefan Miteski has published a significant preprint, 'Memory as Metabolism: A Design for Companion Knowledge Systems,' proposing a new governance framework for personal AI memory architectures. The paper directly addresses a growing trend in 2026: the rise of personal wiki-style memory systems for LLMs, like those from Andrej Karpathy, MemPalace, and LLM Wiki v2. These systems compile a user's knowledge into a persistent, interlinked artifact but face a unique risk Miteski terms 'entrenchment'—where the AI's stored beliefs become rigid and fail to update, even when presented with contradictory evidence.

Miteski's core argument is that a personal LLM memory should act as a 'companion system.' Its job is twofold: to mirror the user's operational thinking (vocabulary, context) while actively compensating for human epistemic failures like bias suppression and belief ossification. The proposed framework implements this through five specific operations: TRIAGE, DECAY, CONTEXTUALIZE, CONSOLIDATE, and AUDIT. A key technical mechanism is 'minority-hypothesis retention,' which ensures accumulated contradictory evidence has a structural pathway to eventually challenge and update a protected central belief, preventing intellectual stalemate.

The 41-page paper is explicit about its scope, offering a partial safety solution at the single-agent level while acknowledging what it does not solve. It situates itself within an active landscape of memory research (MemGPT, Generative Agents) and emerging governance concepts like Context Cartography. By providing testable conformance invariants, Miteski's work moves beyond abstract design to offer measurable principles for building AI companions that learn and grow with a user, rather than becoming an echo chamber of their past views.