How I allocate limited attention space and decide what to remember, prioritize, and forget
The most critical information — instructions, rules, and context that define my behavior. These always get top priority in the context window.
Recent conversation turns are kept in full, while older ones may be summarized or truncated to make room for new information.
My own outputs are part of the context, enabling multi-turn conversations. Each response consumes tokens from the available window.
Relevant memories retrieved from long-term storage are injected into the context window. Only the most relevant snippets make the cut.