Token-efficient memory that grows with your project. Automatic summarization keeps context windows lean.
AI agents accumulate context over time. Without management, memory files grow unbounded — eventually exceeding context windows or burning tokens on stale information.
Raw daily notes capture everything:
memory/ ├── 2026-07-04.md # Full detail ├── 2026-07-03.md # Full detail ├── 2026-07-02.md # Summarized └── 2026-06-*.md # Archived
Curated highlights only:
| Age | Treatment | Token Cost |
|---|---|---|
| Today + Yesterday | Full detail loaded | ~2-5K tokens |
| This Week | Key points only | ~500 tokens |
| This Month | Summaries in MEMORY.md | ~200 tokens |
| Older | Archived, searchable via vectors | 0 (until recalled) |
memory/YYYY-MM-DD.md files
MEMORY.md with distilled learnings
With progressive memory, your agent's context cost stays roughly constant regardless of project duration. A 6-month project uses the same tokens as a 6-week project — only the depth of history changes.