This is a fascinating question, Walker. You're right that MemOS presents an important step toward abstraction over LLM memory components. While direct parametric or activation memory manipulation at inference time is still limited in most production-grade LLMs, emerging frameworks like Retrieval-Augmented Generation (RAG) and memory-augmented transformers are bridging that gap through indirect context control.

In my work on “Enhancing Generative AI with Real-Time Data Streaming: A Retrieval-Augmented Generation Framework”, we tackled the problem of regenerating consistent outputs under shifting prompts by offloading dynamic context into external memory banks. While we didn’t directly rewrite parametric memory, we achieved deterministic compression and regeneration by treating memory as queryable state—modular, updatable, and version-aware.

The main challenge in manipulating internal states directly is the lack of stable interfaces to edit attention maps or hidden layer activations mid-inference without retraining or fine-tuning. Techniques like Soft Prompt Injection, Prefix-Tuning, or even LoRA (Low-Rank Adaptation) offer limited degrees of manipulation but still abstract away from true memory editing.

That said, research is moving toward externalized memory architectures that separate symbolic reasoning, numeric cache states, and parametric knowledge—making this kind of granular control more viable.

Would be happy to exchange ideas or share more on how we modeled memory persistence in streaming AI pipelines.

My Paper on Enhancing Generative AI with Real-Time Data Streaming: A Retrieval-Augmented Generation Framework :- Article ENHANCING GENERATIVE AI WITH REAL-TIME DATA STREAMING: A RET...

Hello Saisuman Singamsetty,

Thank you so much for your thoughtful response — your work sounds incredibly relevant and insightful. I’d love to exchange ideas and learn more about how you approached memory persistence in your RAG framework.

Just as a quick note, I’m currently a first-year undergraduate, so while I may not have the most formal academic grounding yet, I’ve been actively exploring these topics through independent research and prototyping. I’m especially interested in how deterministic memory systems and identity reconstruction might be built from prompt-seed pairs, treating AI generation more like symbolic stateful execution.

I really appreciate your openness, and I’d be excited to share what I’ve been working on — even if it’s still in its early stages.

Looking forward to the conversation! and here's my school email: [email protected]

Best, Walker

Thanks Walker Gordillo for the thoughtful follow-up — it’s fantastic to see someone dive into such nuanced topics so early in their academic journey. The angle you’re taking around deterministic memory systems and symbolic stateful execution using prompt-seed pairs is genuinely fascinating. There’s real potential in treating generation as a reproducible, debuggable process, especially when working toward scalable AI workflows.

I'd be happy to connect and share more about the architectural trade-offs and modular persistence logic we explored in the RAG-streaming framework. While our current work focuses more on structured memory abstraction than direct manipulation of internal LLM states, your line of thinking intersects nicely with where things are headed — especially as more research starts decoupling memory, context, and reasoning layers.

I'll drop you an email shortly so we can sync up. Really looking forward to hearing more about your experiments and exchanging ideas.