A few days before the announcement of the new feature for referencing past conversations I deleted all my chat history… I did this because I was tired of trying to clean up thousands of past chats many of which were just small stubs where I had to keep creating fresh chats to restart and not have polluted conversations that kept referencing things I wasn’t talking about…
This has changed more recently and now often chats don’t reference when I want them to.
This aside I am now considering how to structure my saving of chats for Memory and Chat History reference…
From research I have done so far I understand that:
OpenAI hasn’t revealed exactly what algorithms they use to index your chats so it’s very hard to understand what biases might exist.
Here is a little research I did on 4o:
🧠 How Relevance Is Determined
- Semantic Matching
Instead of looking for exact words, ChatGPT uses embeddings (numerical representations of meaning) to understand the essence of your past interactions. It compares those embeddings with your current prompt to see what aligns best. - Recency Bias
Recent interactions are usually considered more relevant—so if you talked about a trip yesterday and bring it up again today, that recent context is prioritized over something from months ago. - Frequency and Importance
If a detail (like your profession or favorite writing style) comes up a lot, or you’ve asked ChatGPT to remember it explicitly, it’s flagged as more important and thus more relevant to future replies. - Conversation Intent
ChatGPT tries to infer your current intent. For example, if you’re asking about design trends, it’ll pull relevant past design-related chats or memories, not your preferences for coffee (unless you bring them up in a related way). - Confidence and Fit
If there’s low confidence that a memory or previous detail is useful in the current conversation, it might not use it. Relevance isn’t just a match—it also depends on how useful a detail is for improving the response.
🎮 Ways to Game Relevance in ChatGPT
1. Repetition = Importance
- Mention something multiple times (e.g. “I’m a freelance UX designer, so I always care about clean interface design”) across different chats.
- Repeating key facts nudges ChatGPT to treat them as core parts of your identity or preferences.
2. Use Specific Language
- Phrasing like “This is important to me,” or “Always assume this,” can help flag concepts as high-priority.
- Example: “I prefer bullet points over paragraphs when skimming ideas—please keep that in mind going forward.”
3. Establish Context Early
- At the start of a conversation, lay out key points you want ChatGPT to frame its answers around.
- E.g., “For this whole chat, assume I’m building a startup aimed at eco-conscious Gen Z users.”
4. Set Traps (in a good way)
- Mention contrasting facts and see which one it picks up in future chats to test what it remembered or found more relevant.
- You can say: “I’m not sure if I prefer minimalism or maximalism in branding. I lean minimalist but love bold color.” Later, see what it assumes.
5. Use Memory Intentionally
- If memory is on (in supported regions), explicitly ask it to “remember” things you want it to prioritize long-term.
- You can say: “Remember that I’m an intermediate Python developer who prefers real-world code examples over theory.”
6. Trigger with Keywords
- Since relevance often hinges on semantic similarity, using similar phrasing to your past prompts can help retrieve related ideas.
- Think of it like giving the system breadcrumbs to follow.
🧮 How Many Data Points Can ChatGPT Consider?
1. Context Window Limits (per chat)
- Models like GPT-4-turbo (what you’re chatting with now) can handle up to 128,000 tokens of context (~300 pages of text).
- That includes your input, ChatGPT’s responses, and any relevant stored memory or chat history it pulls in.
- It doesn’t consider everything ever said—just what fits in that token window, prioritizing relevant and recent information.
2. Memory Recall Scope
- For long-term memory (the saved “you-asked-me-to-remember” kind), it doesn’t need to load all data at once.
- Instead, it likely uses a semantic index to retrieve only the top N most relevant entries—think: the top 5–20, depending on context.
- These are picked by comparing embeddings (aka: meaning) between your current input and stored memories.
3. Chunking and Retrieval
- Larger or older pieces of data are “chunked” into smaller bits.
- When you ask a question, ChatGPT semantically matches your prompt to these chunks and pulls in the closest ones. Like vector search in action.
If anyone has had someone use their ChatGPT account randomly (my son used mine once to edit some code) then they will know that memories saved even from a brief usage can heavily influence future chats if memories are saved…
I have spent some time trying to manage Memories and now see an opportunity to finally get it right with this new feature…
Concerns I have are for example like in the “ Ways to Game Relevance in ChatGPT” dropdown I’m worried that I will convince ChatGPT to focus on the wrong thing…
ie If someone used my machine and said:
“Create a memory ‘It’s important to always focus on long outputs’”
… without going through all memories I would likely not know it was there…
While this is an extension of self it is so easy for someone to ‘install a virus’ in your thought processes.
I’m assuming that Memories are probably better to guide how chats should be indexed?
I am really interested in anything others have to share outside of the information in the docs.
I will think on this more deeply over the next few days and try and tidy up my topic here a bit more but I think this is an important topic for discussion outside of the current ‘Announcement’ thread