Skip to main content

​
Overview

The Pulse architecture is described in detail in our design paper: “Proactive Cognition in AI Agent Systems via Hierarchical Signal Perception” by Maciej (aniolowie). This page summarizes the key insights and design rationale. For the complete technical specification, see the full paper.

​
The Core Problem

Every AI agent system built today is fundamentally reactive. An agent acts because something triggered it—a user message, a cron job, a webhook, an API call. Remove the trigger, and the agent does nothing. The obvious solution—running the agent continuously, polling for relevant events—would cost approximately $10 per minute at current API prices for capable models, making it economically absurd for any real-world deployment.

​
The Biological Insight

The human brain solves this exact problem. A person doesn’t think about their homework every second of every day. But when they see a familiar file, or a classmate mentions an assignment, something in their brain fires: “I should check on that.” This happens before conscious reasoning. It is fast, cheap, and usually accurate. The brain achieves this through hierarchical perceptual processing:
  • Raw sensory input is processed by cheap, fast, low-level systems that detect change and pattern
  • Only when these low-level systems flag something as potentially relevant does the expensive, slow, high-level reasoning system engage
  • The expensive system (prefrontal cortex) never sees raw sensory data—it only sees pre-filtered, pre-interpreted signals from the layers below it
The Pulse applies this principle to AI agent systems.

​
The Three-Layer Architecture

ENVIRONMENT (files, memory, network, time)
        │
        ▼
  LAYER 1: RETINA          — deterministic change detection, ~0 cost
        │
        ▼
  LAYER 2: LIMBIC FILTER   — tiny neural networks, <5ms on CPU
        │
        ▼
  LAYER 3: PREFRONTAL      — template-based question formation, ~0 cost
        │
        ▼
  AGENT (wakes with focused question)
Each layer acts as a gate. The vast majority of environmental changes are filtered out by Layer 1 (no delta detected) or Layer 2 (low relevance score). The agent is only woken when all three layers agree that something is worth attention.

​
Layer 1: The Retina

Deterministic event detector watching file system events, memory namespace changes, time signals, and optionally network events. It has no state and no intelligence—it only detects and emits structured SignalEvent objects.

​
Layer 2: The Limbic Filter

The core learning component. It maintains a small neural network (LSTM or TCN with ~50,000-200,000 parameters) for each registered module cluster. These models take a sliding window of recent signal events and produce a relevance score (0.0-1.0). The critical innovation for cold-start: module fingerprints—structured JSON documents that each module provides at registration, describing what signals are associated with its relevance. These fingerprints are converted into synthetic training examples that pre-initialize model weights before any real data exists.

​
Layer 3: The Prefrontal Filter

Converts a relevance signal into a specific, actionable question for the agent through template interpolation. For example: “A new file appeared at /home/user/Downloads/hw3.pdf. Is this file related to a course assignment?” The agent wakes up with this specific question, not a blank slate.

​
Key Properties

​
Zero LLM Cost in Normal Operation

The Pulse runs entirely on local compute. Layer 1 is deterministic. Layer 2 is a tiny neural network running on CPU. Layer 3 is string interpolation. No API calls are made. The cost of running the Pulse continuously is electricity, not tokens.

​
Improving Accuracy Over Time

Unlike a cron job, which is equally accurate (or inaccurate) forever, the Pulse improves with usage. Each agent activation provides a training signal. Over time, the cluster models learn the user’s specific patterns.

​
Privacy by Design

All data—training examples, model weights, signal history—is stored locally on the user’s machine. Nothing is sent to external servers. The Pulse does not require an internet connection to function.

​
Design Philosophy

AI is used only where no deterministic process can do the job. Everything else is infrastructure.
The Pulse embodies this principle recursively: it achieves proactive behavior using zero LLM calls in normal operation. The LLM is only engaged when all three layers cannot resolve ambiguity—which should be rare.

​
Read the Full Paper

The complete design paper includes:
  • Detailed comparison with existing approaches
  • Mathematical foundations of the LSTM/TCN models
  • Cold-start strategy and module fingerprint specification
  • Analysis of limitations and future work
  • References to neuroscience and sequence modeling literature
Read the full paper: WHITEPAPER.md

Architecture Guide

Implementation specification and technical details

Module Fingerprints

How modules describe their signal patterns

Build docs developers (and LLMs) love

Get started for freeTalk to us