Capability Injection Without Retraining

Previous attempts to extend the operational capabilities of large language models through context-based knowledge delivery have produced inconsistent results. Systems either fail to adopt injected capabilities reliably, treat injected knowledge as external reference material rather than actionable self-knowledge, or lose the information entirely when context windows are exceeded. We identify two previously unrecognized factors as the primary determinants of injection success: epistemic framing, the structural positioning of knowledge within a trusted first-person memory space; and authorship fidelity, the degree to which injected content matches the model’s own generative voice and style. Using the STAR (Structured Tree with Active Retrieval) persistent memory framework deployed on a 4-billion parameter Gemma 4 (E4B variant) model running locally on a consumer smartphone, we demonstrate that capabilities entirely absent from a model’s training data can be reliably injected, adopted autonomously, and applied without explicit prompting. The model demonstrated unprompted use of a custom device control capability from a vague natural language instruction, and correctly understood and described the intended behavior of a web search capability despite incomplete pipeline execution. These findings suggest that the boundary between trained and injected knowledge is substantially more permeable than previously assumed, and that structured persistent memory systems represent a practical, infrastructure-free alternative to fine-tuning for capability extension across models of any scale.