Retain best practices for coding/chat agent integration that has forking/compaction

[noctuid]

Based on the best practices and #510, the ideal way to retain a chat is to keep retaining the full chat. The hermes integration retains individual turns though.

Is this for cost reasons or because of forking/compaction or something else?

Ignoring cost, would it be better to give as much of the chat session as possible a single document id? There are different possibilities for doing this while still handling forking/compacting:

Forking:

• Use a new document id when forking and only include content after fork to avoid duplication; tag with session and parent session

Compaction alternatives:

• Use a new document id after compaction and start fresh to avoid duplication; tag with session and parent id
• OR If the harness supports getting all the old messages after compaction, always retain the full chat
• OR Pull the document, append, retain again (not sure about performance but shouldn't matter if in background)

Questions:

• Do you have any suggestions here?
• There is currently no way to append to a document, right?

[nicoloboschi]

Hey, we just fixed the behaviour in the Hermes agent.
In 0.5.0 we also shipped update_mode=append so you can just send to hindsight the new turn

[iRonin]

Thanks, that's great! I am also working on a hindsight package for pi coding agent. One thing I'm wondering about is if it would be better not worry about retaining every x turns due to LLM usage. Is there currently any way to have hindsight store but not immediately process the data? Was thinking it might be useful to wait to process until after I'm done with the session or overnight.

@noctuid any progress with pi integration?

[noctuid]

@iRonin I have a working, opinionated (but pretty customizable) extension that I've been using. It's not shareable yet (needs more review, testing, and updates for the breaking session_switch change pi made recently), maybe in the next week or so hopefully.

[iRonin]

@noctuid here is what my clanker vibed today https://github.com/iRonin/pi-less-shitty/tree/main/packages/hindsight if it has any value to you
Some overlap with https://github.com/anh-chu/pi-hindsight

[noctuid]

Thanks! It looks like that other pi-hindsight is missing some things from mine, e.g. I have queueing to disk, configurable message stripping, more rich automatic tagging, and also allow ingesting old sessions with fork detection to prevent duplication. It is also doing some strange things like injecting memories directly into the transcript (and then removing them to prevent double retention), using hardcoded truncation, and not using the typescript library for whatever reason. I might steal the #nomem idea, but I don't know if I need that.

[iRonin]

@nicoloboschi Hindsight is the best thing since ChatGPT lol my agents have been cured from amnesia for once.