ResearchMarch 29, 2026

Context windows are lying to you: why your AI forgets everything between sessions

hours per week professionals waste re-explaining context to AI tools

Source: Microsoft Research / Salesforce AI→

TL;DR

Context windows grew from 4K to 1M+ tokens but model performance drops 39% in multi-turn conversations. The real gap is session memory — AI forgets decisions between conversations, costing professionals 5+ hours per week in re-explanation.

My AI remembers my name. It forgot the entire project brief I explained twenty minutes ago. This is the paradox of modern AI tools: context windows grew from 4K to over 1M tokens in three years, and the experience of using them barely improved. You still start Monday explaining the same architecture decisions you explained Friday. The window got bigger. The amnesia stayed.

Why didn't bigger context windows solve the problem?

Context windows grew from 4K to 1M+ tokens in three years. The problem did not shrink — it changed shape. A 1M-token window can hold your entire codebase. It cannot hold the reasoning behind your decisions, the patterns you rejected last week, or the style preferences you have corrected a dozen times. Token capacity is not memory. It is a buffer that empties when the session ends.

Why is session memory the real gap in AI tools?

Your AI has amnesia between conversations. Every Monday, you start from zero. Every new chat window forgets what the last one decided. The context window handles within-session recall. Nothing handles between-session continuity. Your architecture decisions, your technology choices, your explicit rejections of patterns that did not work — all of this disappears when you close the tab.

Band-aids exist, but they are band-aids

ChatGPT memory, Cursor rules, .cursorrules files — these are attempts to solve session memory. They remember facts: your name, your language, your framework. They do not remember decisions: why you chose Postgres over Mongo, why you rejected the observer pattern last Thursday, why you scoped the MVP to five features instead of eight. Facts without reasoning are trivia, not context.

Maintain a decision log

Start a text file called DECISIONS.md. After every AI session, write down what was decided and why. Before the next session, paste it in as context. It is manual and tedious, but it proves the concept: explicit decision context, carried forward, produces dramatically better AI output. The AI does not need to remember everything. It needs to remember what you decided and why.

How DriftLess solves session memory

DriftLess persists context across sessions and providers structurally. Your decisions, preferences, and scope boundaries are not facts in a memory bank — they are active constraints that shape every AI interaction. Tuesday's reasoning is still there on Friday. Switch from Claude to GPT mid-project, and your context follows. The AI does not need a bigger window. It needs a persistent one.

Stop re-explaining your project every session.

5 sessions free. $0 AI markup. No card required.

Start building free