memory.py

# cognitive_net/memory.py
import torch
import torch.nn as nn
import torch.nn.functional as F
from collections import deque
from typing import Deque, Dict, Any

class CognitiveMemory(nn.Module):
    """Differentiable memory system with biological consolidation mechanisms"""
    def __init__(self, context_size: int, capacity: int = 100):
        super().__init__()
        self.context_size = context_size
        self.capacity = capacity
        self.memory_queue: Deque[Dict[str, Any]] = deque(maxlen=capacity)
        
        # Memory projection layers with adaptive scaling
        self.key_proj = nn.Linear(context_size, 64)
        self.value_proj = nn.Linear(context_size, 64)
        self.importance_decay = nn.Parameter(torch.tensor(0.95))
        
        # Consolidation parameters
        self.consolidation_threshold = 0.7
        self.age_decay = 0.1

    def add_memory(self, context: torch.Tensor, activation: float):
        """Store memory with dynamic importance weighting"""
        importance = torch.sigmoid(torch.tensor(activation * 0.5 + 0.2))
        self.memory_queue.append({
            'context': context.detach().clone(),
            'importance': importance,
            'age': torch.tensor(0.0)
        })

    def consolidate_memories(self):
        """Memory optimization through importance-based pruning"""
        new_queue = deque(maxlen=self.capacity)
        for mem in self.memory_queue:
            mem['importance'] *= self.importance_decay
            mem['age'] += self.age_decay
            if mem['importance'] > 0.2:
                new_queue.append(mem)
        self.memory_queue = new_queue

    def retrieve(self, query: torch.Tensor) -> torch.Tensor:
        """Content-based memory retrieval with attention"""
        if not self.memory_queue:
            return torch.zeros(64, device=query.device)
            
        contexts = torch.stack([m['context'] for m in self.memory_queue])
        keys = self.key_proj(contexts)
        values = self.value_proj(contexts)
        query_proj = self.key_proj(query.unsqueeze(0))
        
        scores = F.softmax(keys @ query_proj.T, dim=0)
        return (scores * values).sum(dim=0)