Create memory.py
Browse files
memory.py
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from collections import deque
|
5 |
+
from typing import Dict, List, Optional, Tuple
|
6 |
+
|
7 |
+
class CognitiveMemory(nn.Module):
|
8 |
+
"""Differentiable memory system with consolidation and retrieval"""
|
9 |
+
def __init__(self, context_size: int, capacity: int = 100):
|
10 |
+
super().__init__()
|
11 |
+
self.context_size = context_size
|
12 |
+
self.capacity = capacity
|
13 |
+
self.memory_queue = deque(maxlen=capacity)
|
14 |
+
|
15 |
+
# Memory importance parameters
|
16 |
+
self.importance_decay = nn.Parameter(torch.tensor(0.95))
|
17 |
+
self.consolidation_threshold = 0.7
|
18 |
+
|
19 |
+
# Memory projection layers
|
20 |
+
self.key_proj = nn.Linear(context_size, 64)
|
21 |
+
self.value_proj = nn.Linear(context_size, 64)
|
22 |
+
|
23 |
+
def add_memory(self, context: torch.Tensor, activation: float):
|
24 |
+
"""Store new memory with adaptive importance"""
|
25 |
+
importance = torch.sigmoid(torch.tensor(activation * 0.5 + 0.2))
|
26 |
+
self.memory_queue.append({
|
27 |
+
'context': context.detach(),
|
28 |
+
'importance': importance,
|
29 |
+
'age': 0.0
|
30 |
+
})
|
31 |
+
|
32 |
+
def consolidate_memories(self):
|
33 |
+
"""Memory consolidation through importance reweighting"""
|
34 |
+
for mem in self.memory_queue:
|
35 |
+
mem['importance'] *= self.importance_decay
|
36 |
+
mem['age'] += 0.1
|
37 |
+
|
38 |
+
# Remove unimportant memories
|
39 |
+
self.memory_queue = deque(
|
40 |
+
[m for m in self.memory_queue if m['importance'] > 0.2],
|
41 |
+
maxlen=self.capacity
|
42 |
+
)
|
43 |
+
|
44 |
+
def retrieve(self, query: torch.Tensor) -> torch.Tensor:
|
45 |
+
"""Attention-based memory retrieval"""
|
46 |
+
if not self.memory_queue:
|
47 |
+
return torch.zeros_like(query)
|
48 |
+
|
49 |
+
keys = torch.stack([self.key_proj(m['context']) for m in self.memory_queue])
|
50 |
+
values = torch.stack([self.value_proj(m['context']) for m in self.memory_queue])
|
51 |
+
query_proj = self.key_proj(query)
|
52 |
+
|
53 |
+
scores = F.softmax(keys @ query_proj.t(), dim=0)
|
54 |
+
return (scores * values).sum(dim=0)
|