Reflexion — Actor, Evaluator, Self-Reflector, Episodic memory
Actor
ReAct trajectory
Evaluator
binary / heuristic / self
Self-Reflector
natural-language diagnosis
Episodic memory
bounded list of reflections
prepended next trial
feed-back loop
three evaluator types
scalar binary/numeric ground truth (tests pass, answer correct)
heuristic stuck-loop, too-many-steps, repeated-action signatures
self-eval LLM-as-judge on own trace (weak; pair with tool grounding)
2026 instantiations
Letta sleep-time compute
a separate reflector agent runs off the hot path, writes to memory blocks
pro-workflow /learn-rule
manual reflection capture into project memory — curated over time
Claude Code /memory
CLAUDE.md functions as long-lived episodic buffer across sessions
LangGraph reflection nodes
graph node that scores and routes to refine-or-finish
memory rot: compact, TTL, or dedupe — accumulating reflections go stale