A memory prosthesis for LLM agents that learns from usage. Not better search — associative memory that improves with every interaction.
Blog (中文): 我让三个 AI 互相竞争进化，两天后它们发明了一个我看不懂的算法 | 知乎
LLM agents are brilliant amnesiacs — every conversation starts from scratch. Existing memory systems (RAG, vector search) are libraries: they store things but always return the same results for the same query.
meomory replaces static retrieval with online-learning associative memory. A 256×256 weight matrix that adjusts after every feedback signal — the more you use it, the better it gets.
The algorithm wasn't written by a human. Three AI models discovered it through evolutionary search.
Embedding: qwen3-embedding 8B → random projection to 256d. The DGD learning layer is a 256×256 weight matrix on top.
Our initial evaluation trained and tested on the same questions. The numbers looked impressive but did not reflect real-world generalization:
| Method | LoCoMo P@1 (1976 Q) | LongMemEval P@1 (470 Q) |
|---|---|---|
| Cosine similarity (no learning) | 22.4% | 41.7% |
| BM25 keyword matching | 30.9% | 48.3% |
| Hand-written DGD | 32.8% | 56.0% |
| AI-evolved algorithm (train=test) | 60.1% | 95.5% |
When we tested on unseen questions (train on 500, test on 1476), the evolved algorithm scored 0.9% — worse than cosine baseline. The 60.1% was a fitness illusion caused by overfitting to the evaluation set.
We redesigned the evaluation: each question is tested before the algorithm learns from it. Questions are shuffled across conversations to expose cross-topic interference.
| Method | Online P@1 (1976 Q, shuffled) |
|---|---|
| Cosine similarity (no learning) | 22.4% |
| AI-evolved algorithm (v6) | 26.0% |
The real gain from online learning is +3.6pp over cosine — modest but genuine. The evolved algorithm (gen107) discovered experience replay, row normalization, and learning rate decay to mitigate catastrophic interference.
- Shared weight matrices cause catastrophic interference: learning query A degrades retrieval for query B
- Fitness function design matters more than algorithm design: train=test rewards memorization, online streaming rewards generalization
- The evolved algorithms are useful for repeated query patterns: same user asking similar questions over time. Not useful for completely unseen queries.
- Evolution independently discovered experience replay, residual connections, and adaptive learning rates — not because it "invented" them, but because selection pressure guided it toward known effective techniques.
Three AI models evolving unsupervised on a server:
qwen3.5:9b (30%) — mass exploration, local model
DeepSeek (20%) — precise optimization
Opus 4.6 (50%) — highest quality, primary driver
v4 (train=test): From 26.8% to 87% in two hours — fast but illusory. Evolved Adam + Momentum + dual-channel + BatchNorm.
v6 (online streaming + shuffled): From 22% to 26% overnight — slow but real. Evolved experience replay + row normalization + residual connections. Completely different algorithm direction under different selection pressure.
As a reranker plugin (plug into any memory system):
from src.mem0_integration.dgd_reranker import DGDReranker
reranker = DGDReranker({"dim": 256})
# Rerank after retrieval
results = your_memory_system.search(query)
reranked = reranker.rerank(query, results, top_k=5)
# Feedback: tell it which memory was correct
reranker.feedback(query, correct_memory)
# Next time, similar queries rank betterRun evolutionary search (discover better algorithms):
python scripts/run_funsearch.py \
--ollama-host http://localhost:11434 \
--deepseek-key sk-xxx \
--iterations 0 --rounds 1 --samples 12 \
--seed-from multi-archsrc/funsearch/ — Evolution framework (island model + multi-model + sandbox + numpy)
src/evolution/ — Prompt evolution (Judge prompt 28.9% → 90.9%)
src/mem0_integration/ — Mem0 reranker plugin
src/dgd.py — Original DGD associative memory
src/bench/ — Benchmark framework (65 experiments, 2446 questions)
scripts/dashboard.py — Real-time evolution dashboard
experiments/ — All experiment results + evolution population data
This project grew from my cognitive science knowledge base (49 concept cards). An AI read my notes on memory, forgetting, and preconscious activation, then auto-linked them to the DGD algorithm from the Hope/Nested Learning paper (NeurIPS 2025). I had reproduced that paper but focused on a different module — the AI found the piece I overlooked.
meomory = me + memory。打错了，留下了。