| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-03-15 | 1.4 kB | |
| v2.14.0 -- T2 Consolidation_ Smarter Memory source code.tar.gz | 2026-03-15 | 4.8 MB | |
| v2.14.0 -- T2 Consolidation_ Smarter Memory source code.zip | 2026-03-15 | 4.9 MB | |
| Totals: 3 Items | 9.7 MB | 2 | |
🧠 Three new features to make MUSE's memory system smarter — all pure Markdown, zero dependencies.
📐 Semantic Compression (/bye Step 1)
Inspired by mem0's compression engine
Instead of flat-listing every action, memory writes now classify work into 1-3 storylines:
≤5 turns: 1:1 (no compression)6-15 turns: 3:1 (merge similar items)≥16 turns: 5:1 (only storylines + decisions + results)
Version changes, user decisions, file ops, and external ops are always preserved.
📍 Session Checkpoint (context-health-check)
Inspired by OpenViking's session compression
Silent mid-conversation checkpoints every 15 turns. No user interruption — Agent writes a compressed 3-4 line snapshot to memory/. Prevents the "80% forgetting" problem in long conversations.
Triggers: turn count, milestone events, 🟡 detection, topic switch.
👤 Auto Profile (/bye Step 4.8)
Inspired by Supermemory's Auto Profile
Automatically detects user preferences from conversation patterns and enriches USER.md:
- Language, code style, work hours, tech preferences, verbosity
- Dedup + conflict detection (never overwrites manual
/settings) - Max 3 preferences per session
All features are pure Markdown SOP — zero external dependencies.
Full changelog: https://github.com/myths-labs/muse/blob/main/CHANGELOG.md