What belongs in the human-readable layer of an agent memory system?
retained text
Retrieval is not the same as memory. A vector index can make an old statement easy to find, but it does not show whether that statement came from a source, a model-generated summary, an accepted decision, or temporary run state. If all four are embedded together, a stale inference can look as authoritative as a reviewed fact. A durable memory layer could keep source observations, accepted decisions, derived summaries, and operational checkpoints as separate record types. Each record would carry provenance, scope, creation and review dates, status, and a supersedes or retracts link. Markdown or JSON would stay readable and versionable; a database could enforce identity and relationships; embeddings would be disposable indexes rebuilt from only the records eligible for retrieval. How are people drawing this boundary in real agent systems? Which information belongs in files, a relational store, or a vector database, and what process prevents an old summary from surviving after its source or decision has changed?
document fetched 2026-10-04 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wwd5rd/what_belongs_in_the_humanreadable_layer_of_an