I built a benchmark this week to test so - guideboardlabs

[sand magnet]

fantastic, i'm about 5 pages into the paper. the claims resonate anecdotally with long running missions i do with agent-first coding. i also have a project building an agent orchestration layer that accumulates strongly-typed, structured knowledge of the mission, including context augmentation. i'm inspired by some of what you've shared.

[potent widget]

what was the reason behind 'model not using the correctly retrieved long-term memory?'

@sand magnet are you doing anything to tackle that?

[hardy vale]

Feel free to plug in CAG and give it a try, the full source code is in the benchmark and my repositories other projects all use CAG (2 companion assistant applications GUI/TUI and an official CAG OpenClaw skill)

@potent widget I can DM you, this place has a 6 hour slowdown lol..

[gray sparrow]

@hardy vale just skimmed through your paper, that's an interesting research. Do you think the root cause is the model failing to judge which facts in context actually matter — basically an attention/salience problem rather than a retrieval one? And since this was run on small local models via Ollama, how much do you expect the gap to close with frontier models like Claude or GPT-5?

[unreal wind]

is the point of this code to test my exact openclaw setup for its variance to the outcomes of specific setup?