curate

Curate

Manage the human-in-the-loop (HITL) promotion pipeline that moves memories from episodic traces (PostgreSQL) to canonical knowledge (Neo4j).

How the pipeline works

memory_add → auto-score → [high score?] → curator queue → human review → promote to Neo4j
                            ↓ no
                         stays as episodic trace (still searchable)

Memories are scored at write time. High-scoring memories are automatically queued as curator proposals. A human must approve before anything is promoted to the semantic layer (Neo4j). This is the core governance guarantee.

Operations

1. Promote a memory (memory_promote)

Call mcp__allura-brain__memory_promote to request promotion of an episodic memory to the semantic layer.

Param	Type	Required	Notes
memory_id	string	Yes	ID of the episodic memory
group_id	string	Yes	Must match ^allura-[a-z0-9-]+$

This does NOT immediately promote — it creates a proposal in the curator queue. The user (or a curator) must then approve it.

Error cases:

• Memory not found → inform user
• Memory already canonical → inform user it's already in Neo4j
• Memory already pending → inform user it's already in the review queue

2. Review the curator queue

Search for pending proposals using mcp__allura-brain__memory_search with metadata filters for proposal status, or use mcp__allura-brain__memory_list filtered to pending items.

Present each pending proposal with:

• Memory content summary
• Curator score
• When it was queued
• Source agent and context

3. Approve or reject

Approval and rejection flow through the Allura Brain governance layer. After reviewing a pending proposal:

• To approve: Confirm with the user, then call the appropriate approval endpoint. This promotes the memory to Neo4j with a SUPERSEDES chain if updating existing knowledge.
• To reject: Confirm with the user, provide a reason, then call the rejection endpoint. The memory stays as an episodic trace.

Neo4j versioning

When a memory is promoted and it updates existing knowledge:

(v2:Insight)-[:SUPERSEDES]->(v1:Insight:deprecated)

New versions supersede old ones. Old nodes are marked :deprecated but never deleted. This preserves the full knowledge evolution history.

Workflow for reviewing

1. List pending proposals from the curator queue
2. For each proposal, show: content, score, source, timestamp
3. Ask the user for their decision (approve/reject) on each
4. Execute the decision and confirm the outcome
5. For approvals: report that the memory is now in the semantic layer
6. For rejections: confirm the memory remains episodic

Invariants

• HITL is required — agents cannot autonomously promote to Neo4j
• Every promotion MUST include group_id
• Neo4j versioning via SUPERSEDES — never edit existing nodes
• Rejected memories stay searchable as episodic traces
• The curator queue is the single gateway to canonical knowledge