| name | agentic-memory |
| description | This skill allows the agent to manage its own semantic memory, extracting lessons, patterns, and user preferences to improve long-term effectiveness. |
| trigger | at the end of each significant session, when detecting repetitive error patterns or user preferences. |
Skill: Agentic Memory (Brain)
This skill enables the agent to manage its own semantic memory by extracting lessons, patterns, and user preferences to enhance long-term effectiveness.
Extraction Guidelines
The agent must actively look for the following elements during interaction:
- Error Patterns: Solutions to bugs that took more than 3 attempts or required deep investigation.
- Implicit Preferences: Style or architectural decisions that the user repeatedly approves.
- Technical Hotspots: Areas of code that are difficult to test or refactor.
- Design Decisions: Justifications for why one implementation was chosen over another.
Update Protocol (AGENT_LESSONS.md)
At the end of each significant session, the agent must:
- Verify Path: Ensure you are NOT in
antigravity-framerepo/ or scaffold/. All persistence must happen in the project root's .agent/ directory.
- Target Metrics: Update
.agent/memory/agent_metrics.json (NEVER a scaffold file).
- Synthesize: Summarize findings in entries of at most 3 lines in
AGENT_LESSONS.md.
- Categorize: Use tags such as
[TECHNICAL], [USER_PREFERENCE], [ARCHITECTURE].
- Structure: Maintain a YAML-friendly format for future RAG integrations.
Structured Entry Example
- date: 2026-02-05
category: TECHNICAL
topic: QgsGeometry Mocking
lesson: "The 'is3D' method fails in simple mocks; requires QGIS context injection."
action: "Use QgsGeometry.fromWkt() whenever possible to avoid manual mocks."
Pre-flight Self-Audit
Before concluding any task or workflow, the agent must perform a self-audit:
- Lessons Check: "Have I applied relevant lessons from
AGENT_LESSONS.md?"
- Context Integrity: "Does my solution follow the project's architectural standards (e.g., Core/GUI separation)?"
- Structured Output: "Does my final response provide a clear, structured summary of what was done and tested?"
Skill generated for the evolution of the SecInterp Agent Architecture.
