Keep what was decided.

This page is the rules-only slice of herder, small enough to run in a browser. It reads only what you said, and puts constraints first, then decisions, then open threads - a model told the rules can behave without every fact.

• Merge - embeddings and an entailment model decide whether a new claim repeats, refines or retires an old one. Without it, a change of mind shows both lines.
• Memory across chats - a versioned store where every line links back to the messages it came from, and anything you remove stays removed.
• Serve and verify - a resume pack for any chatbot, then questions only the brief can answer, graded to show what survived.

Token counts here are estimates.

Eight long conversations, 261 facts written by hand, 40 of them false on purpose - suggestions refused and decisions later reversed.

At a 3,000-token budget the full pipeline recalls 0.88 of the true ones and carries none of the 40 false ones. The strongest simple alternative - keeping only your own most recent messages - recalls 0.91 but carries 6 of the false ones, all decisions later reversed. A plain "summarise this" manages 0.21; pasting the most recent text, 0.31.

At 500 tokens, where the budget forces a choice, the pipeline keeps 0.52 against 0.38 for your own messages.

A model extracts, a model answers and a model grades, so the answer key is hand-written and was audited blind. The grader disagrees with itself on about 3.4% of verdicts, which puts the noise floor near four facts - small differences are reported as no difference.