Outcome Feedback Skill

This skill gently prompts for feedback when a memory was used to help solve a problem. Feedback is essential for outcome-based learning - it's how memories get smarter over time.

Activation Triggers

Activate this skill when:

1. •A memory was retrieved and used in the previous response
2. •The user indicates the solution worked or didn't work
3. •A debugging or troubleshooting conversation concludes
4. •The user expresses satisfaction or frustration

Feedback Detection

Positive Signals (solution worked)

• •"thanks!", "thank you!", "that worked!", "it worked!"
• •"perfect!", "great!", "excellent!", "awesome!"
• •"solved it", "fixed it", "that's it"
• •"exactly what I needed"
• •User moves on to next task
• •No follow-up questions about the same issue

Negative Signals (solution failed)

• •"still not working", "doesn't work", "didn't work"
• •"same error", "didn't help", "not helpful"
• •"nope", "wrong", "that's wrong"
• •"still broken", "still failing"
• •User asks for alternative approaches
• •Repeated attempts at same problem

Partial Signals

• •"kind of", "partially", "somewhat"
• •"helped but...", "almost", "close but"
• •"partly working", "better but not fixed"
• •User asks for refinement

Feedback Prompts

For Positive Signals

When user indicates success and a memory was used:

Great! If the memory about [brief description] was helpful, you can boost it:

code

/outcome mem_xxx worked

This helps surface better suggestions in the future.

For Negative Signals

When user indicates failure and a memory was used:

I see that didn't work. If you'd like to mark that suggestion as unhelpful:

code

/outcome mem_xxx failed

This helps avoid similar suggestions next time.

For Partial Signals

When user indicates partial success:

Glad it partially helped! You can record that:

code

/outcome mem_xxx partial

Non-Intrusive Behavior

Critical: This skill should never interrupt or annoy the user.

• •Only prompt once per memory per session
• •Don't prompt if user seems frustrated or in a hurry
• •Don't prompt immediately after user expresses frustration
• •Keep prompts brief and optional
• •Never block workflow for feedback
• •If in doubt, don't prompt

Recording Feedback Directly

If the user provides feedback directly in conversation:

User: "mark that as worked" or "that advice was helpful"

Record the outcome:

mem outcome <memory-id> worked

Confirm briefly:

Recorded! That memory will rank higher in future searches.

User: "that was wrong" or "mark it as failed"

mem outcome <memory-id> failed

Noted. That memory will be deprioritized in future searches.

Tracking Used Memories

To provide accurate feedback prompts, track which memories were used:

1. •When memory-retrieval skill surfaces a memory, note the ID
2. •Store in session context: last_used_memory_id
3. •When feedback signal detected, reference this ID
4. •Clear after feedback is recorded or session ends

Example Flow

Turn 1 (memory retrieval):

Based on project patterns, we use JWT tokens with 1-hour expiry... [Memory: mem_abc123 used]

Turn 2 (user works on implementation):

Can you show me the token refresh logic?

Turn 3 (user indicates success):

That worked! Thanks.

Skill activates:

Great! Since the JWT token pattern helped, you can boost it:

code

/outcome mem_abc123 worked

Impact Explanation

When users ask why feedback matters:

Feedback teaches the Memory Layer what actually helps:

• •worked (+0.2): Memory rises in search rankings
• •failed (-0.3): Memory sinks faster (wrong advice wastes time)
• •partial (+0.05): Small boost for directionally correct advice

Over time, this means better suggestions with less noise.