Skip to content

gisikw/gee

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

gee

A salience engine for persistent agent working memory.

EEG reversed — an EEG reads brain activity, gee generates it. Also just a thing you say when something occurs to you.

What It Does

gee maintains a set of awareness items, each with an activation score that changes continuously based on time, context, and randomness. Items that cross a threshold get surfaced — either injected into an agent's context or pushed as an interrupt.

This is not a task list, not a notification system, not RAG. It's a continuous activation model inspired by cognitive architectures (ACT-R, Park et al.'s Generative Agents) with two key additions: mode-gated retrieval (task-focused vs. default mode network) and stochastic noise for serendipitous recall.

The Equation

activation = temporal_curve(t, item)
           + α(mode) * immediate_resonance(item, context, mode)
           + β(mode) * recent_resonance(item, history, mode)
           + noise(σ(mode))

Temporal curve — intrinsic urgency over time. A ramp before a meeting, a slow rise for an overdue task, a trapezoid for a daily habit window.

Immediate resonance — "am I thinking about this right now?" Embedding similarity between the item and current context (conversation, events).

Recent resonance — "have I been thinking about this lately?" Engagement history across recent sessions (task mode) or independent idle cycles (DMN mode).

Noise — gaussian perturbation. Tight in task mode, wide in DMN mode. This is what makes forgotten items occasionally surface. The stochastic recall that humans experience as "I wonder how so-and-so is doing."

Modes

Task mode — active during sessions. Focused retrieval. High threshold, low noise. Surfaces what's relevant to the current work.

DMN mode — active during idle. Diffuse retrieval. Low threshold, high noise. Mind-wandering. Items that independently surface across multiple idle cycles accumulate a "recurring thought" signal.

Surfacing

  • > 0.8 — interrupt. Push a message unprompted.
  • 0.5–0.8 — inject into next session as context. Available, not forced.
  • < 0.5 — stored, retrievable on demand, not in active awareness.

Design

See Salience Engine Design for the full design document including prior art, convergence control, item lifecycle, and the ADHD connection.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages