Core Concepts
Emotional Intelligence
Continuous emotional trajectories grounded in cognitive science research.
Overview
MetaMemory models emotions as continuous trajectories rather than discrete labels. Instead of tagging a memory as simply “frustrated,” the system captures the full emotional arc (for example, frustration → insight → satisfaction), preserving the temporal dynamics that make episodic recall effective.
This approach is grounded in Tulving's research on episodic memory, which demonstrates that emotional context is a primary cue for memory retrieval in biological systems.
Emotional Encoding
The encoding pipeline works in four steps:
- Point encoding: each emotion label is mapped to a 128-dimensional base vector and scaled by its reported intensity
- Temporal weighting: more recent emotional states within an episode receive higher weight
- Mean pooling: the weighted vectors are averaged to produce a single 128d representation
- Trajectory features: four scalar features are appended: volatility (σ), trend (ℓ), velocity (v), and range (τ), yielding the final 132d embedding
Trajectory Features
| Feature | Symbol | Description |
|---|---|---|
| Volatility | σ | Standard deviation of emotional intensity over the episode |
| Trend | ℓ | Linear slope: is emotion improving or worsening? |
| Velocity | v | Rate of emotional change between consecutive states |
| Range | τ | Difference between peak and trough intensity |
Emotional Retrieval Channel
The emotional retrieval channel computes trajectory-level similarity rather than point-level matching. This means a query about “overcoming a difficult debugging session” can match memories with similar emotional arcs (struggle → resolution) even if the specific emotions differ.
const results = await engine.search({
query: 'resolved a frustrating production issue',
emotions: [
{ label: 'frustration', intensity: 0.8 },
{ label: 'relief', intensity: 0.9 },
],
channelWeights: { emotional: 1.5 }, // boost emotional channel
});