Session context: 2026-06-18_Mastermind — Dirk surfaced the universal pain: half an hour in, Claude forgets how the conversation started. Lou’s response included the U-shaped memory curve as the underlying mechanism — and the toolkit for working with it rather than against it.
Core Idea
The model reads your entire context on every turn, but it doesn’t weight every part of it equally. Psychologically, it behaves like human memory: strong at the beginning (the setup, the system prompt, the opening instructions), strong at the end (the most recent exchanges), and systematically weaker in the middle (everything that accumulated between).
This is not a bug — it is a consequence of how attention works in transformers. The practical effect is that instructions buried in the middle of a long conversation are the most likely to get dropped, contradicted, or quietly ignored. The model is not being lazy; it is following the shape of its own attention curve.
The implication for conversation design is significant. If you need the model to follow a constraint reliably, it must live at the beginning or the end — not in the middle. A constraint stated in turn 3 of a 30-turn conversation is in the danger zone. A constraint restated in the most recent turn is in the safe zone.
This is also why bad answers compound. When the model gets something wrong and you correct it inline, the failed attempt sits in the middle of the context. The model reads the failed attempt, your correction, and the successful reattempt — and on subsequent turns, it can get confused about which direction is correct. The correction doesn’t erase the failure; it just adds to the evidence the model is trying to synthesize.
Practical Application
Work the edges. Front-load your highest-priority constraints in the opening message. If a constraint matters through the whole conversation, restate it at the end of a long exchange rather than hoping it survives from turn 3.
Rewind, don’t correct. When the model gets it wrong, the safest move is to rewind to just before the query that produced the wrong answer, adjust the query, and regenerate. The failed attempt never enters the context. See Insight - Rewind, Don’t Re-Correct — Keep Failed Attempts Out of the Context Window.
Compact before the middle gets too long. Using /compact (or a handoff skill) at ~50% context fill keeps the middle from accumulating enough noise to degrade the edges. After compaction, the summary lives at the beginning of the new context — back in the safe zone.
Design conversations with the curve in mind. If you’re building a workflow that will run over many turns, structure it so the most critical inputs and decisions land at the start and end. Treat the middle as execution space — not constraint-setting space.
Coaching question: “Where in your typical AI conversation do you state your most important instructions? And where does the model seem to start ignoring them?”
Evolution Across Sessions
The U-shaped memory curve was named explicitly in 2026-06-18_Mastermind as the underlying mechanism behind the rewind/fork/compact toolkit. It hadn’t previously been named as a standalone principle — it appeared as background context in several context-management insights without being foregrounded. This establishes it as the structural reason those techniques work: they’re all methods for keeping critical information out of the forgotten middle.