Thread · HN Continuity

The part I like most is the LLL rule: "the AI reads it; the human writes it." I ended up at the same place from a different direction. I keep a small index that is loaded whole at the start of every session, and the only way to keep it useful was to cap it hard and make it hold pointers, never content. The moment the model was allowed to write into it, it filled with things that looked important and weren't.

Honest question about "never summarize": how do you decide what goes into context on a turn when the time range is wide, say "last month, about the budget"? Chronological order is great for a day, but a month of raw lines won't fit. Is that where the semantic fallback kicks in, or do you just cut at N results?