TanStack AI Compaction Keeps Chats Rolling

This is the right abstraction for long-running AI apps: compact the model’s working context without destroying the user’s history. The architecture is strong, but the release currently has a frustrating packaging mismatch.

• –Compaction runs incrementally before model calls, preventing long chats and agent loops from hitting context limits.
• –Built-in strategies trade fidelity, cost, and simplicity: eviction is free, summarization preserves meaning, and tool-result clearing targets the biggest token sink.
• –Custom and composable strategies let teams tune compaction for coding agents, support bots, or domain-specific workflows.
• –Persistence checkpoints and DevTools events make compaction observable and repeatable across requests.
• –The repository describes a functional 0.1.0 package, while npm currently serves 0.0.0 as setup-only; developers should verify the published version before installing.