Does your AI seem to "forget" as a chat drags on, drifting off track the longer you talk? Usually the issue isn't poor memory; it's that you aren't managing its memory in layers. AI memory isn't a single thing but five distinct layers, from immediate working memory to a long-term external knowledge base. Knowing which data belongs in which layer, when to clear it, and when to hand off is what keeps AI accurate over time. Here is each layer, plus a practical workflow.
AI memory has more than one layer: the 5-layer overview
AI memory splits into five layers: layer one is the immediate Context Window (working memory); layer two is per-conversation summary and handoff; layer three is persistent setup via Projects / Custom Instructions; layer four is account-level global memory; layer five is external memory (files, RAG, cross-platform memory tools). Each differs in capacity, persistence, and purpose — treating them as one is the root of "forgetting."
Layer 1: The Context Window (working memory) and context rot
The Context Window is all the text the AI can currently see — its working memory. As of March 2026, Claude Opus 4.6 and Sonnet 4.6 support a 1 million token window, roughly 750,000 words. But fuller isn't better: as tokens grow, accuracy and recall decline, a phenomenon called "context rot." That makes what you put in more important than how much space you have.
Layer 2: Conversation summary and the 60% handoff rule
The most effective long-chat technique is the "handoff process": when a conversation reaches about 60% capacity, ask the AI to summarize progress, then start a fresh chat pasted with that summary. Push a thread to 90% and reasoning drops sharply while hallucinations rise, as the model loses track of earlier logic. Regular summarizing and handoff is the key move for sustained accuracy.
Layer 3: Projects / Custom Instructions (persistent setup)
This layer holds stable information you use every time. Custom Instructions record your role and preferences for use across all chats; Projects bind background files and a system prompt within a project, callable by every chat inside it. Put long-lived context here so you don't repeatedly paste it into the Context Window, which also eases context rot.
Layer 4: Account global memory (Claude vs ChatGPT)
Global memory means facts remembered automatically across chats, but vendors differ. ChatGPT memory has two parts — Custom Instructions and saved facts — both applying to every chat in the account. Claude scopes memory to a Project; each project holds its own knowledge and Claude does not maintain a global cross-chat memory. Understanding the difference tells you where sensitive information ends up.
Layer 5: External memory (files / RAG / cross-platform tools)
External memory is persistent storage outside the model: uploaded files, a RAG knowledge base, note systems, even cross-platform memory tools. It can preserve your context, preferences, and knowledge across conversations, across days, even across different AIs. When information is large and needs to accumulate over time, this — not the current chat — is your true long-term memory store.
What goes in each layer, and when to clear it
The allocation rule is simple: immediate task content goes in the Context Window and is cleared when done; long-chat progress goes to the summary layer with regular handoffs; stable preferences and project background go to Projects / Custom Instructions; cross-chat facts go to global memory but get reviewed regularly; large long-term knowledge goes to external memory. Remove anything stale or irrelevant from its layer to avoid polluting output.
A practical layered-management workflow
A workflow you can apply: before starting, write stable background into Custom Instructions / Projects; during the chat feed only what this task needs, with key instructions first and key data last; at ~60% capacity, summarize and hand off to a new window; sediment long-term knowledge into an external memory store; periodically prune stale facts from global memory. Follow this and the AI's performance stays stable over time.
Common mistakes
The three most common mistakes: first, dumping all data into the Context Window, triggering context rot and worse answers; second, running one conversation to the end with no handoff, so late-stage logic collapses; third, writing one-off content into global memory, polluting every future chat. Avoid these three and let each of the five layers do its job, and your AI stays accurate and reliable long term. Want to bring layered memory management into your team's workflow? Visit ai.ud.hk to explore UD's AI Staff solutions and see how AI can become a reliable long-term member of your team.
懂AI，更懂你｜UD相伴，AI不冷
