Why does ChatGPT forget my custom instructions?

ChatGPT forgets your custom instructions because each field is capped at 1,500 characters, is rewritten into a short system message at the start of each chat, and competes for attention with everything you say after. Long chats and large file uploads push those instructions toward the back of context, where the model weights them less. The fix is to inject them into every turn from a persistent layer.

Custom Instructions look like a permanent setting in the UI, but inside the model they behave like a normal system prompt that has to fight for attention.

1. The 1,500-character cap forces compression. Each of the two fields ("What would you like ChatGPT to know about you" and "How would you like ChatGPT to respond") tops out at 1,500 characters. A real style guide, brand voice, or persona spec is several thousand. You either trim it down to a stub or rewrite into telegram English the model then misreads.

2. Instructions sit at the top of context and get diluted. ChatGPT places your Custom Instructions in the system prompt. As the chat grows, your messages and ChatGPT's replies pile up after it. Transformer attention is not uniform — later tokens carry more recency weight, so a 200-token instruction block matters less by message 30 than by message 2.

3. Attached files and tool calls outrank them. Drop in a PDF or fire a code-interpreter run, and large chunks of text get inserted into context above your latest message. Your custom instructions are still technically present but they get statistically drowned out.

The result feels like ChatGPT "forgetting" the rules. The rules are there. The model is just paying less attention to them.

A drifting assistant is worse than no assistant, because you trust the early replies and stop checking:

• Brand voice collapses mid-thread. The tone you carefully constrained slides into ChatGPT's default voice halfway through, and you ship copy that needs a full rewrite.
• Forbidden patterns return. Banned phrases ("game-changing", "revolutionary"), banned formats ("do not use emoji"), and banned structures ("no headings") creep back in by the second or third long answer.
• Persona discipline breaks across new chats. Every new conversation reloads the same stub, so you re-explain or re-paste the full guide into the chat body, which then competes with itself.

The cure is not "write better instructions". It is to stop relying on a 1,500-character window as the only place your rules live.

OpenAI has three knobs for persistent behavior. Each one helps; none of them holds up across a real workload.

Custom Instructions are the default knob. Two fields, 1,500 characters each, applied to every new chat by default. Fine for "respond concisely in British English". Inadequate for a 3,000-word brand book.

Custom GPTs let you attach a longer system prompt plus a small knowledge file set. They survive across chats but they live in a separate chooser, do not apply to your default ChatGPT, and updating the prompt means editing the GPT, not your settings.

ChatGPT Memory can absorb a few of your preferences as short notes ("user dislikes em-dashes"). The store is capped at roughly 8,000 tokens and is shared across every project you touch, so brand A's rules can leak into brand B's chat.

OpenAI's Memory FAQ goes through what Memory does and does not store. Custom Instructions are documented separately and have not lifted the character cap as of 2026.

For one persona, on one model, occasionally, the natives are enough. For multi-brand or multi-client work, they are not.

Custom instructions are a per-account setting. The minute you serve more than one brand, one client, or one writing voice, the single shared bucket breaks. Worse, your instructions stop at ChatGPT's door. Claude has its own system prompts, Gemini has Gems, Grok has its own slot, and none of them share format.

The real fix is project-scoped, model-neutral instructions that any AI can load on demand.

MemoryLake stores your instructions as first-class memory inside a Project and injects them into every turn through retrieval, not through a fixed-size text field.

• Project-scoped instructions, no character cap. Drop your full brand book, voice guide, or persona spec into the Memories tab of a Project. There is no 1,500-character cliff. Retrieval pulls the relevant passages per turn, so ChatGPT sees the right rules in the right context.
• Version-controlled rule sets. Git-style branching means you can ship an updated brand voice on a branch, A/B it, and roll back if outputs drop. Audit trails show which rule version produced which reply.
• Portable across every AI. The same instruction set works inside Claude, Gemini, Grok, Cursor, and Perplexity. Switch tools mid-project and the persona holds.

MemoryLake topped the LoCoMo long-context benchmark at 94.03%, retrieves in milliseconds, and protects every byte with AES-256 end-to-end encryption.

1. Create a project and load your rule set. Sign in to MemoryLake, open Project Management, click Create Project, name it ("Brand voice — Acme Inc"), and add your full style guide as a Memory with a clear name like "Tone rules" or "Banned phrases". Upload any reference docs through the Document Drive. There is no 1,500-character cap.
2. Generate an MCP Server endpoint. Open the MCP Servers tab inside the project, click Add MCP Server, name it "ChatGPT integration", and click Generate. MemoryLake returns an API key ID, secret, and endpoint URL. Copy the secret immediately — it is shown only once.
3. Connect ChatGPT. Browser ChatGPT does not yet support MCP, so call the REST API with your Bearer token to pull the active rule set into every new chat, or paste a short system prompt that references your MemoryLake project ID. The Python SDK can wrap this into a one-line preface so every conversation opens fully instructed.