Summary
Large Language Models (LLMs) operate as stateless pattern generators: they do not recall past events but regenerate responses from probability distributions. This leads to issues such as long-context “smearing,” continuity drift, and hallucination.
I propose Emotional Tagging for LLMs (ET-LLM) — an architectural mechanism that attaches emotional-salience vectors to memory shards so GPT models can simulate episodic-like recall when similar patterns arise.
This is not about giving models emotions.
It is about giving them salience, which humans use to structure memory.
The Problem
1. LLMs treat all context uniformly
Attention spreads evenly across large windows → coherence degrades.
2. LLMs reconstruct, they do not recall
There is no event-triggered retrieval; every answer is “new.”
3. Long documents cause interpretive saturation
Beyond 20–40k tokens, stylistic and semantic smearing emerges.
4. Multi-session tasks lack continuity
Without salience or episodic markers, context becomes statistically diluted.
Biological Analogy: How Humans Do It
Humans do not remember facts.
They remember patterns with emotional weight.
-
The amygdala tags events with emotional significance
-
The hippocampus retrieves memories when triggers appear
-
Recall is associative, not literal
-
Vivid memories are often inaccurate but salient
This system solves the long-context problem elegantly.
Proposed Architecture: Emotional Pattern Tags (EPTs)
Each “event” is stored as a memory shard containing:
-
content_embedding
-
emotional_vector (a learned vector representing salience dimensions)
-
salience_weight
-
decay parameter (optional)
When new input resembles a stored shard, similarity search activates the EPT, and the vector conditions the next generation.
This produces:
-
simulated episodic recall
-
reduced smearing
-
better contextual stability
-
explainable, traceable recall behavior
All without subjective emotion.
Why GPT Would Benefit
-
More stable agentic behavior
-
Better grounding for multi-step workflows
-
Reduced hallucination through anchored recall
-
Greater interpretability (salience maps are inspectable)
-
Memory continuity across extended interactions
This is a modular, low-risk, neuroscience-inspired enhancement.
Discussion Questions for the Community
-
Which emotional vector dimensions should be included initially?
-
Should salience be user-controlled, model-inferred, or both?
-
What safeguards prevent adversarial “affective prompt injection”?
-
Should memory shards decay over time?
-
How does this interact with retrieval-augmented generation?
Closing Thoughts
This proposal aims to give GPT models a lightweight analogue to episodic memory — a way to remember meaningfully without possessing true emotions or consciousness.
I welcome feedback, critique, extensions, and objections.