Gavin JuddIndex

Changed my mind · 1 min read · September 2026

More Context Is Not More Truth

The agent often had the information, but couldn't tell which parts were evidence, decisions, proposals, or permission to act.

The central failure of AI memory is not having enough context treating all context as equally true.

What changed

Even with more project history, I kept seeing agents do the wrong work.

They often had the relevant information, but they’d treat a plan as a decision, a summary as a source, or a passing check as proof that the thing actually worked for the user. Sometimes they’d understand the task and assume that meant they had permission to act.

Mostly, the extra context made their wrong answers sound more convincing, which changed what I expected from the memory system.

What I believe now

To decide what to do with the information, a later person or agent needs to know what each item is, where it came from, when it was true, who had authority, and which disagreements are still unresolved. I’d want the memory system to preserve those distinctions.

I think about building memory products differently now too: collecting more history only helps if the next person can tell which parts should affect their decision.

I’m fine with a model interpreting the record, as long as the original is still there to check against.

What survived

Missing context is still a real problem, so it’s worth keeping earlier versions, the sources behind decisions, and details that change how something should be understood.

I no longer expect more history to fix the judgment problem on its own. Remembering that an action came up for discussion doesn’t mean it was approved.

What remains unresolved

How much structure can we add before it becomes a burden to maintain? Which distinctions can the system capture reliably, and which ones need a person to decide?

I’d want preserving those distinctions to be an easy part of the workflow, because otherwise we’re relying on everyone to remember to do it.