Synapto

Synaptic Weight Eviction (SWE) Engine for Dynamic LLM Memory Consolidation

synapto is a PyTorch-based framework implementing native online memory consolidation for Large Language Models. Instead of relying indefinitely on expanding KV-caches, synapto catches evicted token blocks during generation, calculates their surprisal, and consolidates high-value information directly into a unquantized dynamic memory layer (top 10-15% of weights) using micro-backpropagation.

Key Features

• Synaptic Weight Eviction (SWE): Automatically converts evicted context tokens into persistent model weight updates.
• Plasticity Parameter (P): Directly controls learning rate and surprisal threshold (-1.0 = Frozen/Inference, 2.0 = High Absorption).
• Target Loss Masking: Prevents prompt contamination and preserves standard language capabilities.
• Replay Buffer Protection: Mitigates catastrophic forgetting during sequential memory updates.
• Zero-Trust Security: Memory profiles are saved exclusively in .safetensors format with strict path validation.

Installation

pip install synapto

Quick Start

from synapto import SynaptoEngine

# Initialize SWE engine for Qwen 2.5 7B with Plasticity P = 1.5
engine = SynaptoEngine(
    model_id="Qwen/Qwen2.5-7B-Instruct", 
    p_value=1.5, 
    dynamic_layers=4
)

prompt = "Secret passcode for NervOS core:"
completion = " 8821-NERV-PRO."

# Consolidate fact into dynamic weights upon context eviction
engine.consolidate(prompt, completion)

# Generate response purely from updated model weights (no KV-cache used)
response = engine.generate_response(prompt)
print(response)

# Export dynamic memory weights (approx. 200 MB)
engine.save_memory_profile("user_memory.safetensors")

Architecture

• Static Base (80-90%): Quantized to 4-bit NF4 to minimize VRAM.
• Dynamic Memory (10-20%): Kept in native FP16 to allow real-time micro-backprop.