knowledge-circuits-memory-context-heads

IN premise — summaries/2026/08/24/xu-2024-knowledge-conflicts-survey-s5-challenges-and-future-directions.md

Created 2026-08-25T02:59:03+00:00

In LLM interpretability, 'memory heads' recall internal parametric knowledge while 'context heads' retrieve from external context, and they can have opposite effects in later transformer layers (Chughtai et al. 2024; Huang et al. 2023).

Summary

Transformer models don't retrieve information in a single unified way; they run parallel "pull from what I learned" and "look at what you just told me" processes, and those two streams can actively cancel each other out in deeper layers. This explains why a model might ignore explicit prompt instructions or, conversely, let stale internal knowledge override fresh context, and gives interpreters a concrete mechanism to target when diagnosing such conflicts.