RLMs with Differentiable External Memory: Towards Persistent Recursive Reasoning

TL;DR: Imagine giving RLMs a "notebook" they can read from and write to while recursively analyzing huge prompts—can this boost their reasoning and memory? We’ll see if plugging in external, differentiable memory helps RLMs keep track of intermediate results and context across deep recursion trees.

Research Question: Does integrating differentiable external memory architectures into RLMs improve their ability to manage complex, multi-step recursive reasoning over ultra-long prompts?

Hypothesis: Augmenting RLMs with a structured external memory (e.g., as in Lev, 2018; FARM, 2020) will enhance their capacity for persistent information retention across recursive calls, yielding better long-range consistency and deeper reasoning on sprawling documents.

Experiment Plan: Engineer an RLM variant that, at each recursive call, can store/retrieve snippets, intermediate results, or plans in an external memory matrix. Compare to vanilla RLMs and memory-augmented neural nets on tasks like long-form document QA, code reasoning, and multi-hop summarization. Quantitatively assess improvements in answer consistency, error propagation across recursive depth, and retrieval accuracy. Conduct ablations varying memory size, access patterns, and memory-augmented training versus inference-only usage.

References:

• Lev, G. (2018). Neural Networks with External-Memory.
• Challapalle, N., Rampalli, S., Jao, N., Ramanathan, A., Sampson, J., & Vijaykrishnan, N. (2020). FARM: A Flexible Accelerator for Recurrent and Memory Augmented Neural Networks. Journal of Signal Processing Systems.
• Recursive Language Models (2025). arXiv.org.