| // BEFORE/AFTER WALKTHROUGH — See full version in chat or sandbox | |
| // This shows the user experience difference with all 6 improvements | |
| /* | |
| BEFORE: 200+ noisy memories, 3/5 retrieval slots wasted on duplicates | |
| Past context: | |
| - I am allergic to peanuts. | |
| - Nothing with peanuts please ← duplicate | |
| - I hate peanuts ← duplicate | |
| - I like spicy noodles. | |
| - What's a good restaurant near me? ← noise | |
| AFTER: ~80 clean memories, all retrieval slots unique and relevant | |
| About the user: | |
| - Allergic to peanuts | |
| - Likes spicy noodles | |
| - Favorite cuisine is Filipino food | |
| Previous suggestions: | |
| - Suggested Thai place on 5th Ave (already recommended) | |
| RESULT: SLM generates better reply with full context, no repetition. | |
| | Aspect | Before | After | | |
| |---------------------|---------------------------|---------------------------------| | |
| | Memories stored | 200+ (noisy) | ~80 (clean, deduplicated) | | |
| | Context slots | 5 (3 wasted on dupes) | 4-12 (all unique, relevant) | | |
| | Token budget | Fixed | Dynamic (100-500) | | |
| | Prompt structure | Flat list | Typed sections | | |
| | Assistant recall | None | Tracks recommendations | | |
| | Stale handling | Never cleaned | Heat-based eviction | | |
| | Duplicate handling | None | 0.85 threshold auto-dedup | | |
| | Question storage | Stored ❌ | Filtered out ✅ | | |
| | Retrieval | Fixed top-5, 0.45 | Dynamic k=3-12, 0.35-0.50 | | |
| */ | |