LLMs don't have memories of their interactions with humans.

· Bits and Bobs 2/18/25
  • LLMs don't have memories of their interactions with humans.[yr]
    • Another way that the "LLMs are basically a virtual human" mental model is wrong.
    • LLMs have all of the background world knowledge that was statically baked into them during training, but their only "working knowledge" is what's in the context.
      • Their world knowledge is fossilized, frozen in time.
    • The default mode of most chatbots is that each chat is a fresh piece of paper (using only the implied system prompt in the context) to start.
    • In many cases this is convenient; when the LLM starts to "lose the plot" in a certain long-running thread, you can create a fresh one.
      • A pattern I find myself doing a lot for threads that are getting long in the tooth: "Please distill a multi-page executive summary of the main insights and open questions from this thread", and then pasting that summary into a new thread.
    • ChatGPT has started adding features like "memories" but it seems half-baked[ys] and frustrating to use.
      • Some memories I want the system to have are context-independent: fundamental facts about me[yt].
        • Where I live.
        • Books I've read.
        • Concepts I'm familiar with.
      • Some memories are more context-dependent that I don't want to save.
        • That the one open source project I was briefly tinkering with used Deno.
        • That at one point I was asking it questions about my doctor's appointment the next day.
        • That one time when I was trying to fix the carburetor.
          • (Anyone who actually knows me will instantly know this example is a joke… I'm the least handy person in the world.)
      • ChatGPT doesn't distinguish between these situations; it just stores a random subset and then injects them into new threads semi-randomly, which feels confusing and potentially embarrassing.
        • "Why are you bringing up my hemorrhoids in this thread about me trying to understand sparse autoencoders[yu]? Someone could see that!"
      • LLMs seem like they're actively trying to understand you, but it's actually more like talking to a wall, since after the thread is done they forget.

More on this topic

From other episodes