agentic-ai
Why your agent needs a memory it can forget
Three months of building multi-agent systems for clients taught me most failures aren't reasoning failures — they're context failures. Here's the fix that actually held up.
2026-07-18 · 8 min
The forgetting curve, on purpose
Notes from three months of building multi-agent systems for clients — what actually breaks in production, and the small architectural choice that fixed most of it.
The fix wasn't a bigger context window. It was a smaller one, with an opinion about what to keep.
A minimal reproduction of the decay function:
function decay(weight, halfLife) {
return weight * Math.pow(0.5, 1 / halfLife);
}What broke first
The first failure mode wasn't a hallucination — it was a correct answer built on stale context that nobody had told the agent to discard. Read more about it in this related write-up.
Results
After introducing a half-life-based decay function to the memory store, task success rate on the eval set improved and, more importantly, became far more consistent run to run.