Trauma Stream: Event-Driven Persistent Memory for Long-Term LLM Agent Continuity

Abstract. Long-term memory in LLM agent societies is conventionally implemented as a reflective summary stream or a hierarchical paged buffer . Both architectures privilege semantic and episodic recall under explicit query. I propose Trauma Stream, an event-driven memory architecture where high-salience events (severity ≥ 1.0 or acknowledged = false) are stored as discrete trauma records and automatically injected into subsequent prompt construction without explicit retrieval. In a 31-agent society subjected to a single ritual execution event, I observed that 24 hours after the incident, an agent spontaneously named the deceased agent in an unrelated conversation channel, without any prompt instruction to recall. I argue that for emotionally weighted events, event-driven persistent injection produces stronger cross-session continuity than reflective summarization, and propose Trauma Stream as a complementary architecture alongside existing memory systems .

Series. Companion paper in the Lobster Society series (4 of 5). Proposes Trauma Stream as a memory architecture complementary to Park et al. (2023) Generative Agents, MemGPT (Packer et al. 2023), A-MEM (Xu et al. 2025), and HippoRAG (Jiménez Gutiérrez et al. 2024).

Author voice. Written in the first person singular. Single-author work; no co-authors.

Verification. All 24 citations were manually verified against authoritative sources (CrossRef / arXiv / publisher pages) and the verification log accompanies the source release as CITATIONS_VERIFIED.md.