Nova Memory

Project name: Nova Memory

Elevator pitch (tagline): A brain-inspired memory engine for AI agents — it forgets without erasing, exactly like you do. Demoed as "Remember Me", a playable 2D game where the station AI genuinely remembers you across days.

Inspiration

Every AI forgets you the moment the chat ends — and the usual fix is a junk drawer: dump every message into a vector DB and retrieve by similarity. It bloats forever, drowns in trivia, and never forms real understanding. Neuroscience solved this problem millions of years ago: brains encode selectively, consolidate during sleep, and — crucially — forget without erasing. An unused memory goes dormant, yet the right cue brings it back. We built that.

What it does

Nova Memory is a reusable memory engine (Complementary Learning Systems, in code) that gives any agent five-line memory:

Salience gate (dopamine): an LLM pass stores only durable facts — small talk never enters memory
Episodic → semantic consolidation ("sleep"): raw moments merge into a clean user-model; outdated facts are superseded
Biological forgetting: every memory has a strength that decays when unused; salient + rehearsed memories resist. Faded memories go dormant — hidden from recall but NOT deleted — and a strong, specific cue reactivates them ("a smell brings it back"). Recall itself strengthens memories (reconsolidation).
Hybrid recall: semantic + BM25 retrieval over active memories → LLM rerank → a few hundred tokens of context, never the whole history
Memory-grounded generation: the agent can only claim what's in memory — no invented familiarity

The demo, Remember Me, is a playable 2D game: talk to Nova, sleep to consolidate, read her memory of you at the terminal — and step into the Stasis Chamber to skip three weeks and watch minor memories fade to dormant while your allergy persists. Mention your dog, and the faded memory resurfaces, strengthened. python test_bio.py proves the whole cycle deterministically.

How we built it

Everything cognitive and audiovisual runs on Qwen Cloud (Alibaba Cloud Model Studio):

qwen-plus — salience gate, consolidation, reranking, Nova's dialogue
text-embedding-v3 — memory encoding & recall (1024-d)
qwen3-tts-flash — both voices in the demo video (Ethan = narrator, Cherry = Nova)
wan2.2-t2v-plus — every cinematic shot in the demo video

Python + Flask + a single-file HTML5 Canvas game (procedural art & WebAudio — zero binary assets). qwen_client.py is the Alibaba Cloud integration (proof of service usage).

Challenges we ran into

Over-forgetting: our first LoCoMo run scored 10% — the engine forgot too aggressively. We tuned retention, kept episodic detail, and added hybrid retrieval + reranking → 50% vs 42% for a keep-everything baseline (honest: small sample, competitive not superior — hoarding systems win exhaustive-recall benchmarks by design; this track rewards efficiency and timely forgetting).
Cue leakage: dormant memories were resurfacing from generic phrasing overlap. Fixed with an absolute-cosine reactivation threshold — only a genuinely specific cue revives a dormant memory.
Confabulation: with sparse memory, the companion invented shared history (a memory system's worst failure). Fixed with hard memory-grounding in every generation prompt.

Accomplishments we're proud of

Forget ≠ erase working end-to-end and playable live, not just described
9× leaner storage than a store-everything baseline at equal recall accuracy (bench_scale.py)
A demo video whose every visual and voice was generated on Qwen Cloud
Honest benchmarks — we report what we measured, including where bigger stores win

What we learned

Timely forgetting is a feature, not a bug — but it must be retrieval failure, not deletion. And grounding generation in memory is as important as the memory itself.

What's next

Persistence backends (SQLite/Redis), a pip package, LoCoMo at full scale, and dormant-memory "dream replay" during consolidation — dormant memories probabilistically resurfacing during sleep, like real dreams.

Built with: python flask qwen-plus text-embedding-v3 qwen3-tts-flash wan2.2-t2v-plus alibaba-cloud dashscope numpy html5-canvas webaudio