Engrammic

How Engrammic is Different

What sets Engrammic apart from traditional RAG and other memory systems

The Question

Your vector database stores embeddings. Engrammic stores embeddings too. What is different?

The short answer: RAG retrieves documents. Engrammic builds structured knowledge that evolves, explains itself, and detects contradictions.

Side-by-Side Comparison

CapabilityTraditional RAGEngrammic
Cross-session persistenceNone (context window only)Built-in
Provenance trackingNoneEvery node traces to source
Evidence chainsNoneRequired for knowledge claims
Contradiction detectionNoneAutomatic via graph structure
Knowledge evolutionReplace or duplicateSupersession chains (version history)
Multi-hop reasoningLimited to retrieved chunksGraph traversal across relationships
Confidence levelsNoneLayered (observation → fact → belief)
Shared memoryManual syncMultiple agents share one graph

Multi-Hop Reasoning

RAG retrieves chunks that are semantically similar to your query. Engrammic can traverse relationships.

Example scenario:

Your agent stored three things over multiple sessions:

  1. Memory: "Alice mentioned she prefers TypeScript over JavaScript"
  2. Knowledge: "This project uses pnpm workspaces" (evidence: package.json, pnpm-workspace.yaml)
  3. Wisdom: "This team values monorepo tooling" (based on nodes 1 and 2)

Query: "Why do you think this team would adopt Turborepo?"

RAG approach: Searches for chunks mentioning "Turborepo" or "team preferences." Finds nothing directly relevant.

Engrammic approach:

  1. Recalls the belief about monorepo tooling
  2. Traces to supporting evidence (pnpm workspaces, team discussion)
  3. Traverses related nodes (TypeScript preference suggests build tooling matters)
  4. Synthesizes: "Based on their monorepo setup and TypeScript usage, they value fast incremental builds, which Turborepo provides."

The difference: RAG searches. Engrammic reasons over structure.

Supersession, Not Duplication

When knowledge changes, RAG systems either replace the old embedding or create duplicates. Engrammic maintains version chains.

Node v1: "API uses basic auth"
    ↓ superseded_by
Node v2: "API uses OAuth2" (evidence: auth.py refactor)
    ↓ superseded_by
Node v3: "API uses OAuth2 with PKCE" (evidence: security audit)

You can query the current state or ask "how did our understanding evolve?" Both answers exist.

Confidence Layers

Not all knowledge is equal. Engrammic distinguishes:

LayerCommitmentExample
MemoryLow -- raw observation"User said they like tabs"
KnowledgeMedium -- verified with evidence"Project uses 2-space indent (evidence: .editorconfig)"
WisdomHigh -- conclusion from multiple facts"Team prioritizes consistency over personal preference"

When you ask "why do you believe X?", the answer depends on the layer. A memory traces to a conversation. A belief traces through knowledge to evidence.

When to Use What

Use caseBest tool
Search a static document corpusRAG
Remember preferences across sessionsEngrammic
Explain why the agent believes somethingEngrammic
Index a knowledge base for Q&ARAG
Build shared context across multiple agentsEngrammic
Track how understanding evolved over timeEngrammic

They can coexist. Use RAG for your document corpus. Use Engrammic for agent memory and reasoning. The agent can cite RAG results as evidence when storing to Engrammic.

The Core Difference

RAG answers: "What documents mention X?"

Engrammic answers: "What do I know about X, where did I learn it, and how confident am I?"

That is the difference between retrieval and knowledge.

On this page