Context-Aware Neural Nets

Abstract

Moving beyond basic context windows to persistent graph-based memory structures. This paper explores how appending a lightweight vector memory layer to an LLM allows it to approximate genuine long-term human intent understanding.

1. The Illusion of Memory

In standard LLM interactions, memory is merely the concatenation of previous chat history up to a strict token limit. Once the limit is breached, the model suffers catastrophic forgetting of the user's core identity and early instructions.

2. Graph-Based Memory Injections

We present a technique where named entities, user preferences, and emotional states are extracted continuously by a silent observer model and stored in a temporal graph database. During inference, instead of feeding raw chat history, we inject a highly synthesized summary graph into the system prompt framework.

3. Evaluating "Intent Understanding"

By running standard reasoning benchmarks on models with and without our Context-Aware memory injection, we observed a 64% reduction in "frustrating loops" (where a user must re-explain a previously stated constraint).