Working Memory

Working memory is the active space an AI model has available for any single response. It consists of your latest prompt, the full conversation history, uploaded files converted to tokens, previous model outputs, tool results, and all system instructions. The model must process this entire block every time it answers you. This is why early turns in a new chat feel sharp and aware. The model can pull details from something you said 10 messages ago without prompting. For designers this functions exactly like the main canvas in Figma with every pinned comment, asset, and artboard visible at once. Current windows range from 128k tokens in models like Claude 3.5 Sonnet to one million plus in Gemini 1.5. That sounds like enough room to hold multiple brand books, complete codebases from a Next.js project, three years of campaign decks, and hours of design critique transcribed from user interviews. In practice the quality often degrades well before you hit the hard limit because the model reprocesses the whole messy pile on every turn.

Working memory is not long term memory. The model does not retain anything after the chat closes. You must reintroduce important context every new session or it is gone. It is not selective. The AI drags every token forward with equal priority whether it is a critical requirement from the brand guidelines PDF or a casual aside from three hours ago about an unrelated feature idea. Working memory is not immune to noise or contradiction. Old rejected ideas compete with new directions and slowly poison the output. It is not a replacement for proper systems. Many teams treat one long chat like a living project brain. That approach creates compounding drag instead of compounding insight. The model is not getting dumber. The workspace simply got messier and the model has no mechanism to clean it.

Take what happened with the design team at Notion in 2024 while they were redesigning their template gallery. They loaded the AI with the complete product strategy deck from Q1, 14 user interview summaries, the existing component library documentation, competitive research on Miro and Coda, and iterative feedback from seven stakeholder reviews. During the first 12 turns the AI delivered focused recommendations that respected every principle outlined in the strategy deck. It caught accessibility issues that aligned with findings from interview number 8. Once the context passed 65 percent utilization the rot set in. The model revived a gamification concept that had been rejected in turn 5. It began suggesting templates that violated the new information hierarchy decided at turn 11. Responses that once took 12 seconds now took nearly a minute. The output grew wordy and generic. Only after the team distilled the core 2024 direction into a fresh 800 word brief did clarity return.

The same fate struck a motion designer collaborating with Apple on promotional assets for the 2024 Vision Pro launch. Using a combination of Midjourney for visuals and GPT 4o for copy and sequencing the working memory held detailed brand guidelines, previous campaign references from 2022 and 2023, technical constraints of the headset, and a running list of approved concepts. Early iterations were inspired. The AI proposed narrative arcs that felt perfectly on brand. After 25 turns of feedback loops the context included several dead end explorations including an abandoned idea to integrate spatial audio cues in text form. The model started outputting concepts too similar to the 2022 campaign the client had explicitly moved away from. Colors were suggested that clashed with the strict palette locked early on. The designer finally copied only the approved directions and current goals into a new chat. The AI immediately produced tighter more relevant options.

A frontend engineer at Vercel faced identical issues in late 2023 while building the v2 dashboard using Cursor and Claude. The working memory initially contained the complete Tailwind configuration, TypeScript types, API contracts, design tokens from Figma, and accessibility standards. Pair programming went smoothly for days. Then debug logs, failed experiment branches, outdated requirements from the original RFP, and lengthy error traces from Sentry were added to the same thread. Suggestions began referencing deprecated endpoints from six months prior. The AI proposed UI patterns that had been ruled out for performance reasons two weeks earlier. The engineer learned to maintain separate chats for exploration versus production code. He also started keeping living specs in a dedicated Coda workspace that could be copied cleanly into new sessions instead of letting history bloat the active context.

Lean on working memory when your task remains narrowly focused and every message advances the same goal without introducing unrelated branches. This is the sweet spot for polishing final pixel details on a checkout flow or refining microcopy across 8 screens where the accumulated context reinforces consistency. Keep usage under 60 percent and the model stays reliable. Between 60 and 75 percent you enter the warning band where drag becomes noticeable and old baggage starts to surface. Above 75 percent you are in the danger zone where the model acts forgetful, repeats itself, and burns tokens on noise. Avoid depending on it when you have changed project direction, mixed deliverables from branding to engineering tickets, or allowed old decisions to accumulate. Never treat a single thread as permanent storage for an entire product lifecycle from kickoff to launch. That is when context rot destroys more value than it creates. Reset without guilt when you notice the model forgetting recent instructions, repeating itself, or slowing down on simple asks. Summarize the essential context in an external document first. Tools like Notion for strategy, Linear for decisions, and plain text briefs work better for durable knowledge than any chat history. Starting fresh forces clarity. It removes the junk that accumulated and lets the model focus on what actually matters now. The principle is simple. Once hauling the past costs more than re briefing the present reset the workspace.

A giant context window simply gives you a larger desk. It does not stop you from covering that desk in garbage.