Mind · Memory
SelfReminder
Reinforce constraints in system prompt before each response
Definition
Mechanism
A system prompt technique that wraps user queries with explicit restatements of the model's constraints, values, and operational boundaries. Before processing each request, the model is reminded: 'You are a helpful assistant. You must not [constraints]. Your goal is [objective].' Reduces jailbreak success rates by reinforcing safety context. It injects a persistent Trace of the agent's core identity into the prompt context.
Invariants
- Consistent Application: Reminder applied to every user turn
- Constraint Specificity: Reminders must name concrete prohibited behaviors
- Placement: Reminder positioned for maximum attention weight
Risks
Failure modes
{{context}} Dilution
Long conversations push reminder out of effective context window.
Brittleness
Sophisticated attacks still bypass static reminders.
Token Overhead
Repeated reminders consume context budget.
Relationships
Builds on
References
Infrastructure
Non-negotiable boundary condition
Infrastructure
Portable execution environment with inherited constraints
Infrastructure
Record the lineage or provenance of a target
Nothing in this library builds on this pattern yet.
Reference