Memory Stream

Store natural-language experience records scored by recency, importance, and relevance so retrieval packs the most useful memories into context.

A memory stream is a database of natural-language experience records — each with creation time, last-access time, and text — that feeds a generative agent architecture. Base records are observations (self-behavior, other agents, object states); reflections and plans are written back as the same kind of object so one retriever serves all tiers.

Full streams overflow any window. Score candidates after min–max normalizing each factor:

score = α_recency · recency + α_importance · importance + α_relevance · relevance

Recency decays exponentially with time since last retrieval; importance is an LM “poignancy” rating at creation (mundane vs life events); relevance is embedding similarity to a query describing the current situation. Pack top-ranked memories that fit the budget into the prompt under context engineering. Unlike dumping a running summary alone, selective retrieval keeps specific facts available — hybrid retrieval ideas applied to an agent's own life log rather than an external corpus.

---

Sources

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