Context rot

An AI agent works from the text it is given in one prompt. As that text grows with past messages, outdated facts, and unrelated notes, the answers tend to get less accurate. The model has more to sort through and more chances to rely on the wrong detail.

The text itself may still be correct. The problem is the amount and mix of it.

• Answers drift as conversations and documents pile up.
• Old facts can sit beside new ones, and the model may pick the wrong one.
• Simply adding more context does not reliably fix the answer.

• Recall takes a query and returns memories from a person's private memory. You set the size with a level, from low (8,000 tokens, the default) to extra-high (32,000 tokens), or with maxTokens, and cap the count with limit (1 to 200).
• Memory is dated: a changed value with a newer timestamp becomes the current value, and the earlier value stays answerable by date.

Not by itself. A larger window can hold more text, but the model still has to find the right parts. With Harny you choose how much memory a recall returns, using level, maxTokens or limit.