Memory as Metabolism: A Design for Companion Knowledge Systems

The project 'Memory as Metabolism' addresses the 'context window' and 'forgetting' problems in LLMs by proposing a metabolic architecture—where information is not just stored (RAG) but digested and integrated into a persistent knowledge graph. Despite the evocative 'metabolism' metaphor, the project currently lacks any quantitative traction (0 stars, 4 days old) and exists primarily as a design proposal or paper. Its defensibility is near-zero because the core concepts—summarization loops, graph-based memory, and periodic pruning—are already being explored by heavily funded startups (Mem0, Khoj) and academic pioneers (MemGPT, Stanford's Generative Agents). Furthermore, the 'Frontier Risk' is maximum: OpenAI and Apple are best positioned to implement 'metabolic' memory because they possess the OS-level integration required to observe a user's life continuously. Small, independent repos in this space are highly susceptible to being rendered obsolete by a single feature release from a major lab (e.g., ChatGPT's 'Memory' feature or Apple Intelligence's semantic index). The 6-month displacement horizon reflects the aggressive shipping cadence of personal AI memory features by platform owners.