Deep Interactive Reinforcement Learning (Deep IntRL) speeds up training by incorporating external advice, yet conventional approaches apply each piece of advice only once and then discard it. Persistent Advising stores advice in memory, while Broad‑Persistent Advising (BPA) reuses it in similar states, typically identified by K‑means clustering. This work introduces Encoder BPA, which employs a latent encoder to map the state sequence into a latent vector at every time step. Advice is retrieved by comparing latent vectors, followed by a top‑k aggregation and a consistency check to ensure stable reuse. Experiments across complex and partially observable domains show that Encoder BPA markedly improves learning efficiency and advice utilization in the early training phase, outperforming existing Deep IntRL variants.\n\nReview: Encoder BPA’s latent‑based retrieval offers a more adaptable way to reuse advice, advancing the practicality of interactive reinforcement learning in high‑dimensional and noisy settings.