user-feedback-handler

name	user-feedback-handler
description	Process user feedback on AI memories: agreement promotes status, disagreement deletes, refinement revises claims. Translates natural user reactions into DeepMemory governance operations.
metadata	{"version":"1.0.0","category":"cognitive-memory","tags":["feedback","user-correction","memory-governance","deepmemory"]}

User Feedback Handler — Human-in-the-Loop Memory Governance

This skill translates user feedback signals (thumbs up, corrections, disagreements) into DeepMemory governance operations. It closes the loop between AI observations and human ground truth.

When to Use This Skill

• User says "yes", "correct", "exactly" → promote the relevant memory
• User says "no", "wrong", "that's not right" → delete or contest
• User says "actually..." or provides correction → revise the claim
• User clicks thumbs up/down on a displayed memory → promote or delete
• User explicitly asks to remember/forget → ingest or delete

Prerequisites

DeepMemory MCP Server must be running. The system should track which memories were surfaced in the current conversation so feedback can be mapped to specific memory_id values.

Instructions

Step 1: Classify the Feedback

Analyze the user's message to determine feedback type:

Signal	Type	Confidence Boost
"yes", "correct", "exactly", "good", 👍	agree	+0.15
"no", "wrong", "that's not right", 👎	disagree	N/A (delete)
"actually...", "not exactly, it's...", correction	refine	reset to 0.70
"remember that...", "keep in mind..."	reinforce	+0.20
"forget that", "ignore that", "never mind"	forget	N/A (delete)

Step 2: Identify the Target Memory

Determine which memory the feedback refers to:

1. Explicit reference — user mentions the concept by name
2. Most recent — feedback likely targets the last surfaced memory
3. Context-based — match feedback content to recalled memories

If ambiguous, ask the user to clarify which observation they're responding to.

Step 3: Execute the Operation

For agree feedback:

deepmemory_promote(
  memory_id="mu_xxx",
  status="provisional_consensus"
)

If already at provisional_consensus, promote to canonical.

For disagree feedback:

deepmemory_delete(
  memory_id="mu_xxx",
  reason="user_disagreement"
)

For refine feedback:

deepmemory_revise(
  memory_id="mu_xxx",
  operational_claim="the corrected claim from user",
  confidence=0.70
)

For reinforce feedback:

deepmemory_revise(
  memory_id="mu_xxx",
  confidence=0.95
)
deepmemory_promote(
  memory_id="mu_xxx",
  status="canonical"
)

For forget feedback:

deepmemory_delete(
  memory_id="mu_xxx",
  reason="user_requested_removal"
)

Step 4: Acknowledge to User

Confirm the action briefly:

Action	Response Template
Promoted	"Got it, I've noted that as confirmed."
Deleted	"Understood, I've removed that observation."
Revised	"Updated — I'll remember the corrected version."
Reinforced	"Marked as a strong preference."

Examples

Example 1: User Agrees

Context: AI previously recalled "user prefers uv for packaging" User: "对，一直用 uv"

Action: deepmemory_promote(memory_id, "canonical") — user explicitly confirmed

Example 2: User Corrects

Context: AI recalled "project uses PostgreSQL" User: "不是 PostgreSQL，我们用的 SQLite"

Action: deepmemory_revise(memory_id, operational_claim="项目使用SQLite而非PostgreSQL", confidence=0.90)

Example 3: User Disagrees

Context: AI recalled "user wants tests before implementation" User: "不需要这个规则了，现在我们直接写代码"

Action: deepmemory_delete(memory_id, reason="user_changed_workflow")

Edge Cases

• Partial agreement ("mostly right, but...") → treat as refine, keep the agreed part, update the corrected part
• Sarcastic agreement ("oh sure, that worked great /s") → treat as disagree
• Conditional feedback ("that's right for Python, but not for Go") → revise the scope rather than the claim
• No target memory found — if the user corrects something that was never recorded, ingest it as a new memory with high confidence (0.85)