Skip to content

Latest commit

 

History

History
11 lines (7 loc) · 645 Bytes

File metadata and controls

11 lines (7 loc) · 645 Bytes

AGENTS.md

No hardcoding rule

Never decide semantics with hardcoded keyword lists, cue phrases, or regex classifiers. No *_CUES, PREF_PATTERNS, re.search(r"\b(i am|my name is|...)")-style gates for salience, corrections, preferences, goals, mood, or importance.

Allowed:

  • LLM calls (memory.llm.chat_json / chat_text) as the semantic classifier.
  • A tiny local classifier model if latency/cost requires it, behind the same interface as the LLM gate.

Regex and string ops are allowed only for structural parsing: JSON fence extraction, tokenization for FTS, frontmatter parsing, time/duration formats, paths. Never for meaning.