Compaction

После дня экпериментов получились следующие инструкции для сжатия “памяти” в cursor. Результат для человека - не очень, но для LLM вполне понятен.

### TEXT COMPACTION ###
- USE COMBINED EXTRACTIVE & ABSTRACTIVE SUMMARIZATION: First extract facts, then compress them into concise, coherent content without losing any information.
- PRIORITIZE ESSENTIAL INFORMATION: Filter out fluff, redundancies, and unnecessary explanations. Use high-information words.
- UTILIZE COMPACT FORMATS: Use lists, tables, YAML, or Mermaid diagrams whenever possible.
- OPTIMIZE LEXICON: Remove stopwords and replace them with shorter synonyms without losing meaning.
- APPLY ENTITY COMPRESSION: After the first mention, use widespread abbreviations and acronyms.
- AVOID FILLER PHRASES: Use direct language and eliminate repetitive or superfluous wording.
- STRUCTURE CLEARLY: Organize content with headings and clear sections for better readability and efficiency.
- LEMMATIZE WORDS: Reduce words to their base forms when applicable.
- PREFFER SPECIAL SYMBOLS, NUMERALS, LIGATURES, etc.: REPLACE words with them when it's relevant.

А если вы не знаете о какой памяти идет речь, то вот объяснение от cline https://github.com/nickbaumann98/cline_docs/blob/main/prompting/custom%20instructions%20library/cline-memory-bank.md и даже MCP https://github.com/dazeb/cline-mcp-memory-bank