You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have been thinking about how agents keep track of long conversations. The context window is small and gets filled with old messages and long tool outputs. This makes the agent lose focus over time.
I wrote a small thought experiment called The Forgetful Society. In it people remember only three things at a time but they still build complex things by writing notes and breaking work into small steps. This feels like how an LLM agent works. The context window is the forgetful person and the file system is the notes.
From this I sketched a few simple parts. One is an Issue Tree where each message has a path like slash zero slash one so the agent knows exactly where it is in the history. One is lazy tool output where the agent sees only a short tag like at lazy data instead of a long block of text. The real output is stored outside and fetched only when needed. One is a script library where common actions are saved as small files so the agent can reuse them.
These are just design sketches with no tests yet. I share them here because oh my openagent is a large community thinking about agent design. I would like to know if these ideas feel useful or if people here have tried similar things.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
I have been thinking about how agents keep track of long conversations. The context window is small and gets filled with old messages and long tool outputs. This makes the agent lose focus over time.
I wrote a small thought experiment called The Forgetful Society. In it people remember only three things at a time but they still build complex things by writing notes and breaking work into small steps. This feels like how an LLM agent works. The context window is the forgetful person and the file system is the notes.
From this I sketched a few simple parts. One is an Issue Tree where each message has a path like slash zero slash one so the agent knows exactly where it is in the history. One is lazy tool output where the agent sees only a short tag like at lazy data instead of a long block of text. The real output is stored outside and fetched only when needed. One is a script library where common actions are saved as small files so the agent can reuse them.
These are just design sketches with no tests yet. I share them here because oh my openagent is a large community thinking about agent design. I would like to know if these ideas feel useful or if people here have tried similar things.
My notes are at https://github.com/D7x7z49/llm-context-idea. I am happy to hear any thoughts or corrections.
Thank you.
All reactions