Reflection Tree

Periodically synthesize recent memories into higher-level insights that cite evidence, forming trees from raw observations up to abstract self-models.

A reflection tree grows when an agent only has raw observations and fails to generalize (picking the frequently seen neighbor over the collaborator who shares a research passion). Trigger reflection when accumulated importance of recent events crosses a threshold; then: take recent records, ask the LM for salient high-level questions, use those questions as memory stream retrieval queries, extract insights that cite evidence record IDs, and store each reflection with pointers to citations. Reflections may cite other reflections — leaves are observations; higher nodes are more abstract.

This differs from Reflexion's trial-failure verbal RL: the goal is ongoing self-modeling and social inference inside a generative agent architecture, not only fixing a failed task trajectory. Still pair with episodic reflection memory discipline — bound how many abstract nodes enter any one prompt — so reflection does not crowd out the observations that grounded it.

---

Sources

• Generative Agents (Park et al.), pp. 1–22