Working with Memory

This guide covers the common patterns for working with Engrammic: picking layers, setting decay, tagging, linking, and updating.

Picking a layer

You have	Use	Layer
Observation, no source	remember	Memory
Claim with source	learn	Knowledge
Conclusion from facts	believe	Wisdom
Reasoning right now	reason	Intelligence

Setting decay

Memory nodes decay over time. You pick the decay class when you write:

{
  "tool": "remember",
  "arguments": {
    "observation": "Deadline is Friday",
    "decay_class": "ephemeral"
  }
}

Class	Half-life	Use for
ephemeral	7 days	Session context
standard	90 days	Default
durable	540 days	Frequent references
permanent	5 years	Foundational

Tagging

Tags make recall more precise. Add them when you store:

{
  "tool": "remember",
  "arguments": {
    "observation": "Auth uses JWT RS256",
    "tags": ["auth", "security"]
  }
}

Then filter by tag when you query:

{
  "tool": "recall",
  "arguments": {
    "query": "auth",
    "tags": ["security"]
  }
}

Confidence

When you use learn, the confidence score indicates how much to trust the source:

Score	Source quality
0.9+	Official docs, verified
0.7-0.9	Reputable blogs, internal docs
0.5-0.7	Forums, inference
<0.5	Speculation

Linking

You can create explicit relationships between nodes:

{
  "tool": "link",
  "arguments": {
    "from_node": "node_auth",
    "to_node": "node_jwt",
    "relationship": "IMPLEMENTS"
  }
}

Available relationship types: SUPPORTS, CONTRADICTS, CAUSES, PREVENTS, IMPLEMENTS, EXTENDS, RELATES_TO.

Updating

When information changes, supersede rather than delete. This preserves history:

{
  "tool": "learn",
  "arguments": {
    "claim": "Rate limit is 200 req/min",
    "evidence": "https://docs.example.com/changelog",
    "supersedes": "node_old"
  }
}

The old node stays in the graph with an edge to the new one, so time-travel queries still work.