Mind · Reasoning
Reflexion
Self-critique and retry after failure
Definition
Mechanism
A self-improvement loop where after completing a Task attempt, the agent generates linguistic self-critique analyzing what went wrong with the Plan regarding the Goal, then uses this reflection as context for a retry. Unlike fine-tuning, the model is frozen—improvement comes from explicit textual feedback stored in an episodic memory buffer. It writes self-critique to a Scratchpad memory buffer, using this feedback to improve the next attempt.
Preconditions
- Failed trial
- Outcome signal
Postconditions
- Improved policy/next trial
Invariants
- Memory of error persists
- Self-correction based on past failure
Risks
Failure modes
Shallow Reflection
Critique is generic ('try harder') without actionable insight.
Reflection Drift
Each retry diverges further from the goal.
Premature Success
Declaring success before the task is actually solved.
Relationships
Builds on
Accepts
Infrastructure
The atomic unit of intent
References
Qualitative feedback generation
Infrastructure
Specification of desired end state
Infrastructure
Realized state after an action, captured for Plan-vs-reality comparison
Infrastructure
An ordered sequence of steps to achieve a goal
Working memory for intermediate reasoning steps
Used by
Composes with
Step-by-step reasoning (Macro for Think(Chain))
Society
High-velocity solver requiring parallel runtime
Solver implementing the five-surface contract (polymorphic across substrates)
References
Use your own tool before shipping
Recursive self-verification of reasoning
Reference