Hopfield + Hebbian hybrid memory system for LLMs. Two nights of experiments (16 iterations), validated on LongMemEval (ICLR 2025). Architecture: - Single-hop: Two-Stage Hopfield (NN top-20 → softmax settle) - Multi-hop: Hebbian W matrix with WTA pattern separation - 64% on LongMemEval (500 questions), retrieval-only, no LLM dependency - 4ms latency @ 20K memories, ~1GB VRAM Key findings: - Hopfield attention solved noise tolerance (20% → 100% vs flat Hebbian) - WTA pattern separation enables 20K+ capacity - Multi-hop associative chains (6 hops, CosSim=1.0) — RAG can't do this - MiniLM-L6 is optimal (discrimination gap > absolute similarity) - Paraphrase cue augmentation: 55% → 100% on synthetic, 36% → 64% on benchmark - SNN encoder viable (CosSim 0.99) but not needed for current architecture
63 lines
2.3 KiB
Markdown
63 lines
2.3 KiB
Markdown
# LongMemEval Benchmark 结果
|
||
|
||
## 数据集
|
||
|
||
LongMemEval (ICLR 2025, MIT License): 500 个问题,6 种类型,真实多轮多 session 对话。
|
||
|
||
## 结果
|
||
|
||
### Retrieval-only(最终方案)
|
||
|
||
| 类型 | v1 (旧提取) | v2 (改进提取) | 提升 |
|
||
|------|------------|-------------|------|
|
||
| single-session-user | 81% | **86%** | +5 |
|
||
| single-session-assistant | 25% | **82%** | **+57** |
|
||
| knowledge-update | 53% | **71%** | +18 |
|
||
| multi-session | 23% | **53%** | +30 |
|
||
| temporal-reasoning | 29% | **61%** | +32 |
|
||
| preference | 0% | **27%** | +27 |
|
||
| **Overall** | **36%** | **64%** | **+28** |
|
||
|
||
### 加 Gemma 4 推理反而更差
|
||
|
||
| | Retrieval-only | + Gemma 4 |
|
||
|--|---------------|-----------|
|
||
| Overall | **64%** | 40% |
|
||
|
||
Gemma 太保守,检索到了信息但说 "Not mentioned"。不值得增加 1.7s/query 的延迟。
|
||
|
||
## 关键改进(v1 → v2)
|
||
|
||
1. **不截断 assistant 回复**:分段存储(500 字/段)→ single-session-assistant 25% → 82%
|
||
2. **用户自述作为记忆**:用户说的每句话都存一份 → multi-session +30pp
|
||
3. **偏好提取**:正则匹配 "I like/prefer/use/enjoy" → preference 0% → 27%
|
||
4. **日期元数据**:存储 session 日期 → temporal 辅助
|
||
|
||
## 性能
|
||
|
||
- 56ms/query(embedding + Hopfield recall)
|
||
- 平均 22 条记忆/问题
|
||
- 无外部 LLM 依赖
|
||
|
||
## 各类型分析
|
||
|
||
### 强项
|
||
- **single-session-user (86%)**: 用户明确说的信息 → 直接存直接检索,天然适配
|
||
- **single-session-assistant (82%)**: 分段存储解决了长回复截断问题
|
||
|
||
### 中等
|
||
- **knowledge-update (71%)**: 新旧信息都检索到了,top-1 通常是新值
|
||
- **temporal-reasoning (61%)**: 日期信息在 context 里,但检索不做日期计算
|
||
- **multi-session (53%)**: 需要跨 session 聚合,top-K 能召回部分但不完整
|
||
|
||
### 弱项
|
||
- **preference (27%)**: 偏好是隐含的,正则提取覆盖有限。需要 LLM 提取或更多规则
|
||
|
||
## 对比定位
|
||
|
||
64% 在 LongMemEval 上是一个 **competitive retrieval baseline**。论文中的 RAG 基线通常在 40-60%,SOTA(带 LLM 推理)在 70-80%。我们的 retrieval-only 64% 已经超过了多数 RAG 基线。
|
||
|
||
## 结论
|
||
|
||
**Retrieval-only 是正确选择。** 简单、快速、无依赖。提升空间在提取策略(更好的 memory 切分和偏好识别),不在检索架构。
|