Long Context v2.0.0
创建者: 思捷娅科技 (SJYKJ)/zhaog100
MIT License
Copyright (c) 2026 思捷娅科技 (SJYKJ)
免费使用、修改和重新分发时,需注明出处。
Strategy: Startup Optimization
File Hierarchy
| File | Size | Purpose | Loaded When |
|---|---|---|---|
| MEMORY-LITE.md | ~2KB | Core identity, workflows, tracked PRs | Every session startup |
| MEMORY.md | ~10KB | Full history, detailed logs | Only when needed via memory_search |
| memory/YYYY-MM-DD.md | Variable | Daily raw notes | Only when searching specific dates |
How It Works
- Startup: Agent loads MEMORY-LITE.md (not full MEMORY.md)
- Deep recall: Use
memory_searchto find specific information - Detailed read: Use
memory_getto fetch specific lines from MEMORY.md
Token Savings
- Before: MEMORY.md ~28KB → ~7000 tokens
- After: MEMORY-LITE.md ~2KB → ~500 tokens
- Savings: ~6500 tokens per session startup
When to Use
- Session starts with large context files
- Need to recall historical events without loading full memory
- Want to reduce initial token consumption
References
- Compaction patterns:
references/compaction.md - Memory search tips:
references/search-tips.md