🧬 Embeddings · Muscle-Memory Layer

The Sub-Logic Shaping Layer

"The tangent (dμ), not the point (μ)." — embeddings capture trajectory, not position. Models act as STAGE not consciousness.

The Four-Model Psyche

id
JEPA
Gestalt, embodied, fast. Shape-of-trajectory feedback for the embedding space.
muscle memory
Embeddings
Sub-logic shaping. Retrieval as instinct. Trajectory-aware (dμ, not μ).
ego
LLM
Verbal narration. Architect of the stage on which the data plays.
superego
JEV
Principled decision. Verifies muscle-memory matches. Decides retention.
Provider: Dims: 1024 Endpoint: /api/embeddings/*
📈 Trajectory
🔍 Similarity
🧠 Agent Memory
🎯 Curator

📈 Trajectory Explorer (the dμ)

A trajectory = a sequence of states. The endpoint computes the tangent (dμ) between consecutive states and the rolling mean. This is the embedding-as-muscle-memory in action: the model captures the SHAPE OF GOING, not the position.

🔍 Similarity Search

Rank candidate texts by cosine similarity to a query embedding. Use this to find related canon / cells / agents.

🧠 Agent Memory (JEPA-shaped retention)

Each agent state is embedded and assigned a retention strength by JEV. Muscle memory = patterns the agent will recognize again.

🎯 Curator (JEV-verified muscle memory)

Embed a state, rank candidates by similarity, then JEV verifies whether the top match is genuine muscle memory or shallow pattern.

ai-writings theory docs JEV cross-system