Design of Durable Cross-Session Agent State

Identify and specify the design of a durable agent state layer for Claude Code that is neither static instruction files (CLAUDE.md) nor an append-only session transcript, including its data model and lifecycle.

Background

Claude Code today pairs a four-level CLAUDE.md hierarchy for instructions with append-only JSONL transcripts for session histories. The authors argue that there is a missing persistence layer between these two: durable state that carries learning, preferences, or relationship context across sessions without being tied to any single session or static instruction file.

Determining what belongs in this middle layer is central to enabling longer-term memory and collaboration patterns while preserving transparency and auditability.

References

What belongs between these two layers (durable state that is neither a static instruction nor a single session's transcript) is an open design question.

— Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems  (2604.14228 - Liu et al., 14 Apr 2026) in Section 12.2 (Persistence: Memory and Longitudinal Colleague Relationships)