Targeted Memory Reactivation for LLMs — a benchmark of cued vs. uniform replay during consolidation.
Your brain doesn't consolidate memories while you study — it does it while you sleep, and a cue (a sound, a smell) replayed during sleep selectively strengthens whatever it was paired with. An LLM has no such mechanism. dreamcue gives one a "sleep cycle" — a replay pass during downtime — and measures whether selectively replaying user-flagged interactions makes the model retain them better than replaying everything equally.
At an equal replay budget, does cued replay (oversampling flagged interactions) beat uniform replay on retention of flagged facts under interference?
This is a benchmark, not a method. The deliverable is a measured answer.
| Component | Choice |
|---|---|
| Model | meta-llama/Llama-3.2-1B-Instruct (LoRA adapters, FP16) |
| Compute | Modal H100 (NVIDIA / CUDA) — see docs/decisions.md for the AMD-vs-H100 call |
| Dataset | Synthetic key-value associations (~600 facts, 20% flagged, paraphrased probes, ~2000 interference facts) |
| Arms | No-replay, Uniform replay, Cued replay (≥3 seeds each) |
| Phase | What | Gate |
|---|---|---|
| 0 | Environment + repo bootstrap, Modal smoke test | Single LoRA step on MI300X, projected cost < $25 |
| 1 | Dataset + forgetting floor | GATE: flagged probe accuracy drops ≥25 abs pts under interference |
| 2 | Uniform + cued arms | Equal replay budgets verified (token-level), ≥3 seeds |
| 3 | Analysis + artifact | Retention-curve figure + summary table + RESULTS.md |
See docs/ for decisions, tuning notes, and the original PRD framing.
uv sync
modal token new # one-time auth
modal run modal_app.py::smoke_testFull sweep instructions land once Phase 1 gate is met.
MIT. See LICENSE.