engram-offload

name	engram-offload
description	Context offloading: save large outputs to Engram to keep messages lean. Use during long builds or when context > 40%.

Engram Context Offload Protocol

When to Offload

Save to Engram instead of leaving in messages:

• File read > ~100 lines → save key facts (not raw content)
• Large tool output (build logs, test results, search results > 50 lines)
• Discovery you'll need later this session
• Context > 40%

Naming Convention

title:     {project}/{slug}/{type}
topic_key: {project}/{slug}/{type}

Examples:

• my-app/auth/schema — auth module schema
• my-app/session/state — state before compact
• my-app/build/errors — current build errors

Saving

mem_save(
  title: "{project}/{slug}/{type}",
  topic_key: "{project}/{slug}/{type}",
  type: "discovery",
  content: "{compressed key facts — not raw output}"
)

Save what you learned, not what you read.

Retrieving

Previews are always truncated — always two steps:

Step 1: mem_search("{project}/{slug}") → get ID
Step 2: mem_get_observation(id) → full content

Compact Protocol

Before /compact:

mem_save(
  title: "{project}/session/state",
  topic_key: "{project}/session/state",
  type: "discovery",
  content: "task: ...\nfiles_read: [...]\nkey_facts: [...]\nnext_steps: [...]"
)

After /compact:

mem_search("{project}/session/state") → mem_get_observation → continue