Neural Procedural Memory: Using Implicit Activation Steering to Improve LLM Agent Memory Without Training

Summary

This paper introduces Neural Procedural Memory (NPM), a training-free framework for LLM agents that replaces explicit textual instructions (like RAG) with implicit activation steering. Instead of injecting symbolic guidelines into model contexts, NPM distills procedural skills from historical contrastive experiences into steering vectors in the activation space, directly activating task-relevant neural mechanisms. Evaluations across four agent benchmarks show NPM performs comparably to explicit instruction baselines, and combining implicit steering with explicit workflows yields complementary benefits. Representational analyses reveal that the steering vectors encode consistent task logic with organized structures in the activation space.

Source

Twitter / XNeural Procedural Memory: Using Implicit Activation Steering to Improve LLM Agent Memory Without Trainingarxiv.org