I've been using Jev in my coding agents for compaction and task classification. I want to modify the compaction so that it is reversible.
The problem with compaction or any kind of context management is that you have to decide what to keep before you know what you'll need in the future. A giant error log may not seem useful when you're trying to solve one problem, but then when you move on to a related problem the error log might help a lot. If you've compacted that part of the history away, you may not be able to get the information back.
Inspired by @VictorTaelin's OptMem, here's what I want to try.
> Keep the messages and tool results in an append-only log, each with an event ID. Compact would still shrink the context. The full log would stay outside it.
> Use OptMem's tree to group older events under summaries with pointers back to the originals. Have Jev rank those summaries against whatever the agent is doing now, then open the relevant branches.
> Set a token limit for retrieval, but stop earlier if it finds what's needed. Return the actual log fragments through a tool result, with their IDs.
fast-jev-compaction already uses Jev to filter tool history. I want to try that same judgment step for deciding what to bring back.
Havenβt built this part yet, but Iβm gonna give it a shot
@shedsgns Meanwhile I plugged jev into my coding agents for context compaction and task classification. It picks what survives compact and adjusts thinking based on task complexity. Bro is doing housekeeping in my setup xD
Btw tested some other local sultions
@MonidHQ Installed monid for some short simple task and then forgot about it. But then found out in the logs that the agent had been using it to generate some images and for 90% of the research it did. Lol