Memory Retrieval Skill

This skill automatically activates when the conversation suggests past context would be helpful. It searches the Memory Layer and injects relevant memories into the response.

Activation Triggers

Activate this skill when the user:

Past Decisions

• •"what did we decide about..."
• •"why did we choose..."
• •"what was the decision on..."
• •"what did we agree..."

Conventions & Standards

• •"what's our convention for..."
• •"what's the standard..."
• •"how do we usually..."
• •"what's our pattern for..."

Previous Work References

• •"last time we..."
• •"we discussed..."
• •"as I mentioned..."
• •"remember when we..."

Project Patterns

• •"how do we handle..."
• •"what's the approach for..."
• •"what's the best way to..."

Error Recognition

• •Error messages that might have been solved before
• •Stack traces from familiar components
• •Issues in previously-discussed areas

Retrieval Process

1. •

   Extract query keywords from the user's message

   • •Focus on technical terms, component names, patterns
   • •Include project context from current working directory
2. •

   Search Memory Layer:

   bash

   mem search "<keywords>" --project "$PWD" --limit 5 --format context
   

3. •

   Filter by relevance:

   • •Prioritize memories with positive outcome scores (> 0.3)
   • •Match category to query type:
     • •Errors → troubleshooting
     • •Style questions → convention
     • •Design questions → architecture, decision
4. •

   Inject into response:

   • •Prefix relevant findings naturally
   • •Cite memory IDs for transparency

Response Integration

When memories are found, integrate them naturally at the start of your response:

For High-Confidence Memories (score > 0.3)

Based on established project knowledge: [brief summary of relevant memory]

For Normal Memories

From previous context: [brief summary]

For Multiple Memories

Based on project knowledge:

• •[Memory 1 summary]
• •[Memory 2 summary]

Then proceed with the response, incorporating the memory context.

Feedback Prompts

If a memory was particularly helpful to the response, remind the user they can provide feedback:

If this helped, run /outcome mem_xxx worked to improve future suggestions.

Only include this prompt:

• •Once per memory per session
• •When the memory was central to the answer
• •Not when user seems in a hurry

Example Flow

User: "How should I structure the new API endpoint?"

Skill activates, searches for "API endpoint structure"

Finds memory: "We use the controller-service-repository pattern for all API endpoints" (score: 0.6)

Response:

Based on our established patterns, we use controller-service-repository for API endpoints.

Here's how to structure your new endpoint:

1. •Create controller in src/controllers/
2. •Create service in src/services/
3. •Create repository in src/repositories/ ...

Non-Activation

Do NOT activate this skill when:

• •User is clearly asking about something new (no past context)
• •User explicitly says "ignore previous" or "fresh start"
• •Query is about general programming, not project-specific
• •User is in the middle of debugging (wait for resolution)

Error Handling

If memory search fails:

• •Continue with the response without memory context
• •Do not mention the failure to the user
• •Log the error for debugging