Memory and JSON format support for Assistants

The memory feature has been announced to ChatGPT. It would be nice to have the same for Assistants. I’m developing an application where assistants should participate in ongoing conversations. To keep them remembering important highlights I have to create summaries and store them in the initial instruction (by updating them). This more or less works but I feel like this new Memory feature does pretty much the same out of the box. Why not add it to assistants?

Another question is about adding JSON output format for Assistant API. I remember back in November 2023, the documentation promised that this feature was coming soon. Is it still the case? I have to be creative with prompts to enforce the JSON output format but it is not very reliable. This feature would be really helpful

1 Like

OpenAI doesn’t track who your customers are using threads to keep such a thing organized.

Assistants has a run parameter “additional instructions”.

Assistants has a tool calling ability that can emit information in accordance with the specification’s schema.

You have an API execution environment where you can have a list of memory facts that have been previously emitted which you can load up into the AI instruction context.

Just connect the dots. Or have extreme levels of control when you write your own code and directly access models per-turn through chat completions.

JSON can be added simply by instructing how the output shall be created, and prohibiting markdown containers (```) in responses. An ultimate aid to guide output is a json schema. If you don’t do this, the “JSON Mode” will fail dramatically anyway.

Thank you for the reply. Yeah, both of your suggestions are exactly what I’m doing now. I understood the Memory feature as something that automatically extracts a conversation summary and stores it somewhere. I thought the same could be applied to threads, but it looks like there is some fundamental difference between threads and a ChatGPT session. As for the JSON output format, I’ve just noticed that the documentation has changed since I checked it the last time. It is recommended always to add some instruction that the response should be provided as a JSON. It is curious what exactly the JSON Mode does in the completion API. But anyway, I got my answers.