Memory: checkpoints and store

Maps to: Memory: checkpoints (short-term), store (long-term).

Scope

Thread-scoped conversation state versus durable memories that survive across threads, users, or organizations.

Design questions

• What belongs in thread checkpoints only versus namespaced long-term store?
• Retention, export, and deletion policies per tenant and per data class.
• Semantic retrieval versus exact keys; embedding refresh when models change.
• How harness code reads and writes memory without coupling to a single vendor schema.

Tradeoffs

• Rich long-term memory improves personalization but complicates compliance and debugging.
• Storing full traces in checkpoints aids replay but increases cost and PII surface.
• Custom backends trade operational burden for storage characteristics (latency, residency).

Evaluation hooks

• Multi-turn thread continuity after process restart.
• New thread sees prior user preferences written in an earlier conversation.
• Migration drill: export store, change model provider, verify retrieval quality.

Reference notes

See LangChain runtime article (memory figure).