Skip to content

Novelty Detection

The Problem

Without surprise detection, an AI agent treats all memories as equally important. A routine "code compiled successfully" gets the same importance as "production database corrupted." This leads to:

  • Important memories drowning in noise
  • Critical errors being forgotten as quickly as routine events
  • No adaptive importance — every memory starts at the same baseline

How It Works

The surprise detector maintains a running statistical model of "normal" memory vectors using Welford's online algorithm (numerically stable one-pass mean/variance). When a new memory arrives, its L2 distance from the running centroid is converted to a Z-score:

flowchart LR
    A["New Memory<br/>L2 norm = 3.7"] --> B["Running Stats<br/>μ=2.1, σ=0.6"]
    B --> C["Z-score<br/>(3.7 - 2.1) / 0.6 = 2.67"]
    C -->|"Z > 2.0"| D["⚡ Surprising!<br/>importance = 0.85"]
    C -->|"Z < 0.5"| E["😐 Normal<br/>importance = 0.4"]

    style D fill:#e74c3c,color:white
    style E fill:#95a5a6,color:white

Dual Importance Formula

Importance is computed from two independent surprise dimensions:

\[\text{importance} = \alpha \cdot \sigma\left(\frac{x - \mu}{\sigma}\right) + \beta \cdot \text{temporalNovelty}\]

Where \(\sigma()\) is the sigmoid function, \(\alpha = 0.6\), \(\beta = 0.4\).

Dimension Signal Weight
Spatial surprise Z-score of L2 norm vs. running statistics 0.6 (default)
Temporal surprise Time since last memory with matching tags 0.4 (default) 
flowchart TD
    VEC["Memory embedding vector"] --> SPATIAL["Spatial Surprise<br/><i>Z-score vs running centroid</i>"]
    TAGS["Synaptic tag pattern"] --> TEMPORAL["Temporal Surprise<br/><i>time since last matching tags</i>"]

    SPATIAL --> FUSE["Fused Importance<br/><b>α × spatial + β × temporal</b>"]
    TEMPORAL --> FUSE

    FUSE --> FLASH{"Z-score ≥ 3.0?"}
    FLASH -->|"No"| NORMAL["Normal importance<br/><i>0.0 – 0.99</i>"]
    FLASH -->|"Yes"| FLASHBULB["⚡ Flashbulb Memory!<br/><b>importance = 1.0, pinned</b>"]

    style FLASHBULB fill:#e74c3c,color:white
    style FUSE fill:#f39c12,color:white

Why Welford's Algorithm?

Naive variance computation (Σ(x-μ)²/n) requires two passes or suffers from catastrophic cancellation with floating-point arithmetic. Welford's algorithm maintains numerical stability with a single pass — critical for an always-running system that processes millions of memories over its lifetime.

Flashbulb Memory — Extreme Surprise

Biological analog: Flashbulb memories are vivid, long-lasting memories formed during moments of extreme surprise or emotional intensity (e.g., hearing about a major world event). The amygdala signals the hippocampus to strengthen encoding.

When the Z-score exceeds a threshold (default: 3.0), the flashbulb policy activates:

Effects:

  • Importance is set to 1.0 (maximum)
  • The pinned flag is set — this memory is exempt from temporal decay in Phase 4 of the scoring engine inside the Recall Pathway
  • The memory will persist indefinitely unless explicitly forget()'d

Use Case

An AI coding agent encounters OutOfMemoryError for the first time (Z-score: 4.2). This triggers flashbulb encoding — the error memory is pinned at maximum importance and will always surface when the agent encounters memory-related issues.

Where It Fits in the Remember Pathway

Surprise detection happens at Step 3 of the Remember Pathway:

flowchart LR
    EMBED["Step 1: Embed text<br/><i>→ vector</i>"] --> TAGS["Step 2: Encode tags<br/><i>→ Bloom filter</i>"]
    TAGS --> SURPRISE["Step 3: Surprise Detection<br/><b>Z-score → importance</b>"]
    SURPRISE --> FLASH["Step 4: Flashbulb Check<br/><i>pin if extreme</i>"]
    FLASH --> DEDUP["Step 5: Deduplication"]
    DEDUP --> WRITE["Step 6: Write to tier"]

    style SURPRISE fill:#f39c12,color:white
    style FLASH fill:#e74c3c,color:white

Next Steps