Mind · Memory
LatentAttachment
Vector embedding bound to a card without entering its hash
Definition
Mechanism
Attaching a high-dimensional vector embedding (latent) to a symbolic pattern card. This allows agents to use 'Fuzzy Search' to find the card, while still using the 'Canonical Hash' for strict verification. It bridges the gap between LLM intuition and code execution. It binds a vector embedding to a ConceptAnchor, enabling semantic search without breaking the symbolic hash.
Preconditions
- Embedding model is fingerprinted/versioned
Postconditions
- Pattern is discoverable via vector search
Invariants
- The latent block is outside the canonical form: removing it leaves the hash unchanged.
- The embedding is derived from the card's own content.
- Records which model produced it.
Risks
Failure modes
Model shift invalidates embeddings.
Relationships
Builds on
References
Infrastructure
Pinning meaning to immutable references
Used by
References
Mapping to known solutions
Infrastructure
Structured capability advertisement: agent_id, endpoint, protocols, capabilities, metadata
Offline exploration of embedding space