Multi-Vector Encoding

Overview

Rather than representing each memory as a single embedding vector, MetaMemory encodes every memory into four specialized vector spaces. Each space captures a different dimension of the memory, enabling richer retrieval.

Semantic Embedding (1536d)

The primary representation. Generated via OpenAI's text-embedding-ada-002 model, capturing the semantic meaning of the memory content. Stored in Pinecone for fast approximate nearest-neighbor search.

Emotional Trajectory (132d)

Captures the emotional dynamics of a memory. The 132 dimensions break down as:

• 128d: mean-pooled base vectors scaled by emotion intensity
• σ (volatility): how much emotional state fluctuates
• ℓ (trend): overall emotional direction (improving or worsening)
• v (velocity): rate of emotional change
• τ (range): spread between emotional extremes

Point encoding maps each discrete emotion label to a base vector, scales it by the reported intensity, then applies temporal weighting before mean-pooling.

Process Sequence (132d)

Encodes the sequence of actions or steps within a memory. Particularly useful for debugging and procedural memories:

• 128d: positional encoding of process steps
• 4 features: sequence length, complexity, branching factor, step regularity

Context Embedding (64d)

Captures the situational context surrounding a memory:

• Task type (debugging, planning, research, etc.)
• Domain (infrastructure, frontend, data, etc.)
• Complexity level
• Time-of-day features
• Additional metadata features

Fusion Weights

The four embedding spaces are combined using learned fusion weights α, trained via gradient descent on retrieval feedback. Default weights:

These weights adapt per-context as the system accumulates retrieval feedback, shifting emphasis toward whichever embedding space yields the best results for a given query type.