Skip to content

Habituation

The Problem

Without habituation, an AI agent repeatedly recalls the same "most relevant" memories — creating a filter bubble. If memory A has the highest similarity score, it dominates every recall, crowding out potentially useful but slightly-less-similar memories.

Query 1: "database issues" → [A, B, C, D, E]     ← A dominates
Query 2: "database issues" → [A, B, C, D, E]     ← Same results!
Query 3: "database issues" → [A, B, C, D, E]     ← Filter bubble

With Habituation

Query 1: "database issues" → [A, B, C, D, E]     ← Fresh results
Query 2: "database issues" → [B, C, A, D, F]     ← A drops, F emerges
Query 3: "database issues" → [C, F, B, G, D]     ← New memories surface

How It Works

The habituation system tracks recall frequency per memory ID and applies an exponentially increasing penalty:

flowchart LR
    RECALL["Memory recalled"] --> TRACK["Track recall count<br/><i>per memory ID</i>"]
    TRACK --> COMPUTE["Compute penalty<br/><b>penalty = rate ^ (count - 1)</b>"]
    COMPUTE --> APPLY["Adjust score<br/><b>score × penalty</b>"]
    APPLY --> RESULT["Reduced score<br/>→ drops in ranking"]

    style RECALL fill:#4a90d9,color:white
    style APPLY fill:#e74c3c,color:white
    style RESULT fill:#f39c12,color:white
\[\text{penalty}(n) = \text{rate}^{(n-1)}\]

Where \(n\) is the number of times this memory has appeared in recall results and \(\text{rate}\) is the configurable decay rate (default: 0.85).

Penalty Curve

Recall # Penalty (rate=0.85) Effect
1st 1.00 Full score
2nd 0.85 15% reduction
3rd 0.72 28% reduction
5th 0.52 Half score
10th 0.20 80% reduction
20th 0.04 Nearly eliminated

Decay Rate Configuration

The default decay rate of 0.85 provides a balance between novelty and relevance. A higher rate (0.95) creates a gentler penalty — useful when the agent genuinely needs to recall the same memory frequently. A lower rate (0.70) aggressively surfaces new content.

Where It Fits in the Pipeline

Habituation is applied after the 6-phase scorer produces results, but before final ranking:

flowchart TD
    SCORER["6-Phase Scorer<br/><i>produces top-K candidates</i>"] --> SUPPRESS["Step 4: Suppression Filter<br/><i>remove explicitly blocked</i>"]
    SUPPRESS --> HAB["Step 5a: Habituation Penalty<br/><i>attenuate repeated results</i>"]
    HAB --> GRAPH["Steps 5c–5e: Graph Augmentation<br/><i>Hebbian, Temporal, Entity</i>"]
    GRAPH --> SORT["Final Sort → Top-K"]

    style HAB fill:#e74c3c,color:white
    style SORT fill:#00b894,color:white

Key: The penalty multiplies the score field — it doesn't modify the underlying memory. Habituation is a recall-time effect, not a storage-time effect.

Interaction with Other Systems

System Interaction
Reconsolidation Habituation reduces recall score, but reconsolidation increases the memory's durability. A frequently-recalled memory resists temporal decay (fewer buckets) but gets a lower score on repeated queries.
Surprise Detection New, surprising memories start with high importance and no habituation penalty — they naturally dominate initial queries.
Suppression If a memory is fully suppressed, habituation is irrelevant — it's excluded at Step 4 before habituation is applied.

Next Steps