echo-hindsight-llm: An Open Toolkit for White-Box Emotional Memory

Raising $20,000active

An open-source library implementing somatic-marker-style emotional memory for open-weight LLMs — activation-level signals from past outcomes that improve model decision-making.

Project Details

In this project, we will develop echo-hindsight-llm, an open source toolkit to give open-weight models an emotional memory, which cognitive scientists have shown is essential for humans to make good judgments.

The capability, framework, and prototype code have all been developed for our recent papers: The Echo Amplifies the Knowledge: Somatic Marker Analogues in Language Models via Emotion Vector Re-Injection and Scaling the Echo: Multi-Scale Validation of Somatic Marker Analogues in Language Models via Emotion Vector Re-Injection with accompanying Substack articles: https://jaredglover.substack.com/p/what-happens-when-you-give-an-ai and https://open.substack.com/pub/jaredglover/p/does-a-gut-feeling-scale.

In those papers, we extracted internal emotional states as activation-level vectors from past experiences and reinjected them during subsequent deliberation phases on related tasks. The mechanism was shown to successfully generalize across the Gemma family with model scales ranging from 1B to 27B parameters, significantly improving its decision making in high risk scenarios (from 50-52% to 72-98% with the biggest benefit at the largest scale).

This project isn’t about making AI more emotional. It’s about making AI judgment work the way human judgment actually works — not through rules alone, but through the accumulated weight of experience.

Theory of Impact

Emotional memory is largely uncharted territory in AI alignment. Deployed systems currently rely almost exclusively on external, text-based guardrails that remain highly vulnerable to adversarial circumvention and prompt injection. In contrast, activation-level emotional memory establishes a robust substrate for autonomous, self-regulating behavior which is firmly rooted in established cognitive science research findings. By embedding a latent signal of past consequences directly within the internal layers of the network, the model can dynamically modulate its present actions based on the valence of historical outcomes. This technical framework shifts alignment away from fragile, external constraints toward an internalized representation of consequences.

Furthermore, this toolkit serves as the essential architectural foundation for downstream safety mechanisms, such as self-triggered honesty loops, where internal distress features automatically flag deceptive policies before text generation occurs. Open-sourcing this pipeline democratizes access to a vital safety dimension that no other group is currently packaging. By delivering a production-grade, open-source repository for Gemma and Llama, we enable independent labs and agent developers to integrate and build upon activation-layer somatic signaling as a standard defense layer for autonomous agents.

People

Jared Glover

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$20,000

Ideal

$70,000

What this ask is

Funding will deliver a concrete, pip-installable library under a permissive MIT/Apache license, accompanied by a reference evaluation harness and reproducible notebooks, with initial support for Gemma-3 (4B/12B/27B) and Llama-3.3-70B. The initiative will be executed by the author of the core method (PhD, EECS, MIT), with the ideal budget adding support for more model families, HuggingFace transformers integration, and LangChain wrappers.

This project isn’t about making AI more emotional. It’s about making AI judgment work the way human judgment actually works — not through rules alone, but through the accumulated weight of experience.

How the money will be spent

Minimum covers ~3 months to package and release the emotional-memory pipeline for Gemma-3 and Llama-3.3 with worked examples and evals. Ideal adds engineering for broader model coverage, a decision-making benchmark suite, HuggingFace transformers integration, and LangChain wrappers.

Discussion

No comments yet. Be the first to share your thoughts.