Retain batch of independent facts or full conversations

[hadleyrich]

Hi there,

I’m loving working with Hindsight so far, awesome job on the ease of use and deployment.

I’m working on an agent that tunes both its personality and learns about its user.

I currently have implemented two banks, one for personality traits and users memories and preferences.

From looking at the retain docs and the simple examples there I initially decided to have my LLM summarise a conversation into a list of distinct facts/knowledge and submit these in a batch retain to Hindsight.

After getting a little more comfortable with how Hindsight works I wonder if I’m trying to poorly duplicate what Hindsight will do itself.

From a Hindsight perspective, do you think it would be better or at least equivalent for me to pass an entire conversation history as one retain call and let hindsight take care of the rest?

If I should be passing just the conversation, would it be better to be provided as e.g. structured JSON or just newline delimited user: assistant: lines?

This is an example of my summarization prompt:

1. Any facts mentioned about users, preferences, or context
2. Any patterns in how the user communicates or what they prefer
3. Any important information that should be remembered for future conversations

Only include information that would be useful to remember long-term. Be concise.

Conversation:
{conversation_text}

Extract and format as a numbered list of facts/preferences (one per line, starting with a number):```

Thank you once again for an excellent product. 

[nicoloboschi]

Hey! Yes Hindsight automatically read and extract fact from the content. The best way to use it is to put the entire conversation in any format you like.

If you already are preprocessing and have to decide a format anyway, Hindsight has some optimization on chunking if the input is JSON (because most conversations are in json format [{"role": "user", "content": ".."}]