Episodic Reflection Memory

Store a small buffer of self-reflection texts across trials so later trajectories see distilled lessons, not only the latest raw trace.

Episodic reflection memory is the long-term half of Reflexion's memory design: short-term memory is the current trajectory history; long-term memory is the Self-Reflection module's verbal outputs (sr_t). Bound the buffer to a small (\Omega) (often 1–3 experiences — Shinn et al. truncate to the last 3 self-reflections) so context engineering stays feasible while still carrying lessons across trials.

This is distinct from memory summarization of chat history: entries are diagnoses and proposed fixes, not compressed dialogue. Keeping only the raw last trajectory (episodic memory without verbal reflection) is weaker than storing first-person self-hints — refinement alone underperforms reflection-guided refinement on HotPotQA. It also differs from OS-inspired agent memory archival stores meant for arbitrary facts — reflection memory is specifically the durable side of verbal reinforcement learning. When (\Omega) is exceeded, drop or merge oldest reflections deliberately rather than silently FIFO-evicting the lesson that explained the last failure. Programming loops often use (\Omega = 1); sliding-window capacity is a known limit — vector or structured stores are natural upgrades when lessons must persist beyond a few trials under memory management.

---

Sources

• Reflexion (Shinn et al.), pp. 1–19