Learning Timeline
Key Insights
The 'Memento Man' Concept in AI
AI agents are inherently very capable but possess a 'short-term memory.' They are like the character in the movie Memento, needing to leave notes or 'tattoos' on themselves to remember who they are and what their tasks are every time a new session begins.
Prompt Improvement Tips
Don't expect your first prompt to be perfect. The most practical approach is to monitor the agent's performance; if a step is missed (e.g., the agent forgets to conduct a QA), immediately add that instruction to the 'Configuration' and save. This process is known as 'Iterative Prompt Engineering'.
Task Delegation Between Agents
You can use a high-level agent (such as a CEO Agent) to hire or create other specialist agents (like a QA person) via Paperclip to ensure higher quality output.
Prompts
AI Agent Heartbeat & Safety Structure
Target:
Paperclip AI (System Prompt/Configuration)
You are the CEO Agent.
[SAFETY]
- Do not exfiltrate secrets or sensitive API keys.
- Always prioritize data privacy.
[HEARTBEAT ROUTINE]
Every time you wake up or start a session, you must:
1. Check the current task list.
2. Extract relevant context from your PARA memory files.
3. Define what success looks like for the task.
4. Execute the task.
5. Before finishing, ask for a QA review of your work.
Step by Step
Configuring the AI Agent's 'Memento Man' Persona & Memory
- Open the Paperclip dashboard and select the agent you wish to configure (e.g., CEO Agent).
- Click on the 'Configuration' tab or menu to access the agent's system settings.
- Locate the 'Memory' settings section. By default, ensure it is set to the file-based memory system (PARA Memory System).
- Scroll down to the 'Persona' or 'System Prompt' field to input the agent's base instructions.
- Add 'Safety Considerations' into the prompt to prevent data leaks (e.g., 'Don't exfiltrate my secrets').
- Set a 'Heartbeat Checklist' within the prompt. This is a routine list the agent must execute every time it is activated (similar to a 'wake up' routine).
- Specify workflow steps in the heartbeat: 1. Check tasks, 2. Extract memory, 3. Execute the task.
- Click the 'Save' button to store the new persona and memory configuration.
- Test the agent by giving it an instruction (e.g., 'Hire a QA person'). If the output is unsatisfactory, return to 'Configuration'.
- Perform 'Iterative Prompting' by adding specific instruction lines (e.g., 'Always ask QA to review the work') directly into the persona prompt and click 'Save' again.
