Beyond the Bucket:
The Four Layers of the
MAGMA Architecture

Standard vector databases are one-dimensional: they measure similarity. Point A is close to Point B. That's all they know. This is useful for search, but it's insufficient for reasoning. When your agent needs to answer "What caused the deployment failure last Tuesday?" — a similarity search returns snippets about deployments and failures. It does not return the causal chain that connects them.

VEKTOR implements the MAGMA framework (Multi-level Attributed Graph Memory Architecture) — organizing your agent's memory into four orthogonal layers that each capture a different dimension of knowledge. Together, they enable what we call Associative Pathfinding: the ability to traverse a logical path through history rather than just retrieve similar fragments.

Standard RAG fails at multi-hop questions. "What language does the developer who likes dark mode use?" returns snippets about "Sarah likes dark mode" and "Project X uses Rust" — but unless they appear in the same paragraph, RAG can't connect them. The answer requires three hops: Sarah → dark mode preference → Project X → Rust.

VEKTOR's graph traversal solves this. The entity layer has a node for "Sarah." From Sarah, semantic edges connect to her preferences. From Sarah, causal edges connect to her project contributions. From the project node, temporal edges lead to the current tech stack. The answer assembles itself through pathfinding — not keyword proximity.

// Multi-hop retrieval using cross-layer pathfinding
const result = await memory.recall(
  "What language does the dark mode developer use?",
  {
    layers: ["semantic", "entity", "causal"],
    hops: 3,
    topK: 5
  }
);
// Traversal: dark mode → Sarah (entity) → Project X (causal) → Rust (semantic)
// Returns: ["Sarah uses Rust", "Sarah prefers dark mode", "Project X stack: Rust"]