Heartbeat Intelligence: Adaptive Learning System

[tyejcoleman]

Overview

The heartbeat system currently evaluates signals with static weights and fixed thresholds. This epic tracks work to make the heartbeat learn from user behavior and connect signals intelligently.

Goal: Move from reactive signal detection to adaptive, context-aware decision-making.

Current system

• 14 signal types with static tierWeight map and fixed confluenceThreshold
• Adaptive intervals (4 multipliers: time-of-day, signal density, velocity, recency)
• 3-layer dedup (fingerprint + entity overlap + content hash)
• Single global responseCooldownMultiplier() based on 7-day response rate
• Escalation tracking via topic_surfacings table
• No learning from which signals users actually care about
• No cross-signal connection via knowledge graph
• No absence detection or re-entry handling

Features

• Signal feedback loop: learn which signals users engage with #10 — Signal feedback loop — Track per-signal-type engagement. Auto-adjust confluence weights based on learned engagement rates. New signal_feedback table.
• Signal fatigue tracking: suppress repeatedly ignored signals #11 — Signal fatigue tracking — Suppress signals that fire repeatedly with no user action. Exponential backoff per signal+memory pair. New signal_fatigue table.
• Absence awareness: detect user gaps and generate re-entry briefings #12 — Absence awareness + re-entry briefing — Detect gaps in user activity (>24h). Bypass normal pipeline and generate a prioritized catchup summary.
• Cross-signal reasoning: connect related signals via knowledge graph #13 — Cross-signal reasoning — Use the knowledge graph to connect related signals. Compound signals (e.g., deadline + silent assignee = blocked deadline) with tier boosting.
• Per-topic response profiling: learn user priorities by topic cluster #14 — Per-topic response profiling — Learn per-topic response patterns using entity type clustering. Per-topic cooldown multipliers. New topic_response_profile table.
• Proactive context linking: surface related past decisions when signals fire #15 — Proactive context linking — When a signal fires, search memory for related past decisions and outcomes. Surface connections the user wouldn't see on their own.

Design principles

• Programmatic first, LLM only where it adds clear value (Absence awareness: detect user gaps and generate re-entry briefings #12 re-entry summary, Proactive context linking: surface related past decisions when signals fire #15 context linking)
• All learning tables use rolling windows (no unbounded growth)
• New behavior is additive (doesn't break existing signal evaluation)
• Each feature is independently shippable
• Minimum sample thresholds before learned behavior kicks in (avoid cold-start noise)

Suggested order

1. Signal feedback loop: learn which signals users engage with #10 + Signal fatigue tracking: suppress repeatedly ignored signals #11 (feedback + fatigue) — highest ROI, zero LLM cost, foundational for later features
2. Absence awareness: detect user gaps and generate re-entry briefings #12 (absence) — standalone, simple detection logic
3. Per-topic response profiling: learn user priorities by topic cluster #14 (per-topic profiling) — builds on Signal feedback loop: learn which signals users engage with #10's feedback data
4. Cross-signal reasoning: connect related signals via knowledge graph #13 (cross-signal) — requires good entity extraction coverage
5. Proactive context linking: surface related past decisions when signals fire #15 (context linking) — premium feature, depends on rich memory history