-
-
Find the memory that changed the agent’s mind.
-
The AI ignored “avoid overnight layovers.”
-
Remove one memory at a time. Only M001 flips the answer.
-
M001: HIGH influence, score 1.00. Quarantine it.
-
After quarantine: Flight A → Flight C.
-
Remove one memory. Replay. See what changes.
⚠️ The problem
Personal AI acts on what it remembers. When it goes wrong — ❓ which memory did it?
💡 What RecallGuard does
- 🔁 Replays the bad decision with Nemotron
- ➖ Removes one memory at a time and replays again (3× each)
- 🏆 Ranks memories by influence
- 🛑 You quarantine the culprit → the next decision changes
🧪 Live result (real Nemotron)
| Condition | Result |
|---|---|
| All memories | ❌ Flight A — overnight layover |
| Without M001 | ✅ Flight C — changed 3/3 ❗ |
| Without M002 / M003 | ⚪ Flight A — no change |
| After quarantining M001 | ✅ Flight C |
🛠️ How we built it
- 🧠 nvidia/nemotron-3-super-120b-a12b on Nebius Token Factory makes every decision and every replay
- ✅ Code only compares answers — it never decides the result
- ⚙️ FastAPI · Next.js · SQLite
🚧 Challenges
- ⚠️ Nemotron followed a soft memory → we tuned the seed wording on the live model
- ⚠️ Temperature 0 isn’t fully deterministic → 3 runs per case + majority vote
📚 What we learned
Memory wording steers an agent. Replays need stability checks.
🚀 What’s next
Pairwise removal · more scenarios · factor-level influence
🔗 Live demo: https://recallguard.jxdtw.com · 💻 Code: https://github.com/seantw0301/RecallGuard
Built With
- fastapi
- nebius-token-factory
- next.js
- nvidia-nemotron
- python
- sqlite
- tailwindcss
- typescript
