Memory Packs

Memory packs package selected Dense-Mem knowledge into portable JSON. Use them to import or export memories such as workflows, project conventions, operational patterns, and their supporting evidence between teams or profiles.

For exact tool schemas and artifact fields, see Technical Reference.

When To Use A Memory Pack

Use case	Example
Reuse a workflow	Export a blog-writing workflow learned from project history.
Share operational knowledge	Move a deployment checklist from one profile to another.
Preserve support evidence	Export the facts plus supporting claims and fragments.
Review an external pack	Inspect a pack before importing anything.

Do not use memory packs for secrets, credentials, or personal data that should not move between profiles.

Workflow

Step	Tool	Writes memory?
Find candidates	`find_memory_pack_candidates`	No
Export selected items	`export_memory_pack`	No
Inspect a pack	`inspect_memory_pack`	No
Import a pack	`import_memory_pack`	Yes
Roll back an import	`rollback_memory_pack_import`	Yes

The older *_skill_pack* tool names remain accepted as hidden compatibility aliases, and legacy dense-mem.skill_pack.v1 artifacts can still be parsed. New clients should use the memory-pack names.

inspect_memory_pack is the no-write review path. import_memory_pack creates an import ledger entry and can create or update claims and facts depending on mode, selected items, and conflict decisions.

Find Candidates

json

{
  "query": "blog writing workflow",
  "limit": 10
}

Candidates come from existing active facts and validated claims with supported memory-pack predicates.

Export A Pack

json

{
  "name": "blog-writing",
  "description": "Reusable blog-writing workflow",
  "fact_ids": ["fact_123"],
  "include_support": true
}

The export result includes:

Field	Meaning
`canonical_json`	The portable JSON artifact.
`sha256`	Integrity hash for the canonical artifact.
`filename`	Suggested `.memory-pack.json` filename.

Keep include_support=true unless you intentionally need to omit source evidence.

Inspect Before Importing

json

{
  "artifact_json": "{...}",
  "expected_sha256": "..."
}

Inspection reports duplicates, conflicts, superseded matches, matching claims, and any decisions required before a trusted import can proceed.

For a remote artifact, use an HTTPS URL:

json

{
  "url": "https://example.com/blog-writing.memory-pack.json",
  "expected_sha256": "..."
}

Import Modes

Mode	Behavior
`review`	Imports selected items as reviewable claims. It does not promote source facts to active facts.
`trusted`	Treats the pack as trusted source material and can validate claims, promote source facts, or supersede local facts when decisions allow it.

Trusted URL imports require expected_sha256.

Example review import:

json

{
  "artifact_json": "{...}",
  "mode": "review"
}

Example trusted import with a conflict decision:

json

{
  "artifact_json": "{...}",
  "expected_sha256": "...",
  "mode": "trusted",
  "conflict_decisions": [
    { "index": 0, "action": "demote_to_claim" }
  ]
}

Supported conflict actions are:

Action	Meaning
`import_anyway`	Import despite the local match or conflict.
`skip`	Do not import that item.
`supersede_local`	Replace conflicting local active facts.
`demote_to_claim`	Keep source-fact material as a claim instead of an active fact.

Roll Back An Import

json

{
  "import_id": "import_123"
}

Rollback uses the import ledger. It aborts when affected graph entities changed after import, because reverting those changes could erase unrelated later work.

Agent Skill Export Prompt

Memory packs are for importing and exporting memory artifacts. Dense-Mem also ships MCP prompts for file-based sharing when an assistant should turn selected experience into an Agent Skill instead of creating a memory-pack artifact.

Use prompts/list to discover bundled prompts and prompts/get to render one. The first bundled prompt is export_memory_as_agent_skill; it asks the LLM to recall Dense-Mem context for a topic and draft a shareable SKILL.md file plus optional reference files. The generated skill file is not automatically imported back into memory.