tiered-memory

Apply a tiered memory model to reduce agent context load, organizing data into hot, cold, and wiki tiers to eliminate payload bloat during long-running tasks. This approach prevents agents from loading full history and obsolete data on every spawn.

Manage AI agent memory across sessions by using a cold tier with a 30-day TTL that summarizes cross-session history. This loads on demand to maintain continuity without replaying past interactions, keeping active payloads small.

Tiered memory integrates directly into spawn templates by defining policies where hot memory is always loaded for immediate actions, cold memory is loaded on demand, and wiki knowledge is available when domain expertise is required.

Yes, you can use tiered memory for debugging recurring agent issues. The wiki tier provides durable, structured knowledge of stable conventions, while the cold tier maintains summarized cross-session history to track recurring problems efficiently.

Your Squad agent loads large 74KB payloads on every spawn because it loads its full history, including obsolete data. Implementing a tiered memory model separates hot, cold, and wiki tiers to deliver substantial context reductions.

Tiered memory is highly effective for ADR-driven architecture work because the wiki tier stores durable, structured knowledge for stable conventions. It ensures domain expertise is available on demand without bloating the active agent context.