Core Patterns

This section explains the common decision patterns used in agentic systems with Memintel. These patterns represent reusable ways to convert meaning into deterministic decisions — implemented through Concept → Condition, expressed via a strategy type and parameter definitions.

---

Threshold Detection

Purpose: Trigger an action when a value crosses a fixed threshold.

strategy:
  type: threshold
  params:
    value: 0.8

Use cases: Churn risk detection, fraud probability, system load thresholds.

Replaces non-deterministic checks like if "risk seems high" with a deterministic rule. Simple, interpretable, and consistent.

---

Change Detection

Purpose: Detect significant changes over time.

strategy:
  type: change
  params:
    percentage: 0.3

Use cases: Sudden drop in engagement, spike in errors, rapid change in metrics.

Useful when agents monitor evolving systems. Captures dynamics — detects shifts, not just absolute levels.

---

Percentile-Based Detection

Purpose: Evaluate a value relative to a population.

strategy:
  type: percentile
  params:
    value: 95

Use cases: Top-performing users, outlier detection, ranking-based triggers.

Useful when absolute thresholds are not meaningful. Provides relative evaluation that adapts to distribution changes.

---

Z-Score Anomaly Detection

Purpose: Detect statistical anomalies relative to a baseline.

strategy:
  type: z_score
  params:
    threshold: 2

Use cases: Anomaly detection, unusual system behavior, abnormal financial activity.

Statistically grounded and robust to noise. Allows agents to react to unexpected changes.

---

Trend Detection

Purpose: Identify consistent upward or downward movement.

strategy:
  type: change
  params:
    direction: "up"
    window: "7d"

Use cases: Growth trends, declining engagement, performance degradation.

Captures direction — not just magnitude. Helps agents respond to gradual changes.

---

Divergence Detection

Purpose: Detect mismatch between two signals.

strategy:
  type: composite
  params:
    expression: "narrative_signal - price_signal > threshold"

Use cases: Narrative vs behavior mismatch, expectation vs reality gaps, model vs actual divergence.

Critical for systems where LLM perception might differ from real-world data. Surfaces hidden risk.

---

Composite Conditions

Purpose: Combine multiple conditions into a single decision.

strategy:
  type: composite
  params:
    expression: "(high_risk AND high_value) OR critical_event"

Use cases: Multi-factor decisions, prioritization logic, complex workflows.

Allows structured decision logic without relying on LLM reasoning. Composable, expressive — and still fully deterministic.

---

Pattern Selection Guide

Signal type	Recommended pattern
Stable absolute metrics	threshold
Dynamic / evolving systems	change
Relative ranking needed	percentile
Statistical anomalies	z_score
Categorical matching	equals
Multiple factors combined	composite

---

Combining Patterns

Patterns can be layered for higher precision:

strategy:
  type: composite
  params:
    expression: "z_score > 2 AND change > 0.2"

This allows higher precision, reduced false positives, and better overall decision quality.

---

Key Principles

1. Always separate computation (concept) from evaluation (condition)
2. Express all decisions using strategies and parameters
3. Choose the simplest pattern that works
4. Prefer deterministic strategies over heuristic reasoning
5. Never embed decision logic inside agent prompts