tiered-memory

Reduce agent context payload by implementing a three-tier memory model that separates hot, cold, and wiki knowledge. Agents load immediate hot context by default while keeping durable knowledge stored, preventing bloated payloads and noisy data during long-running tasks.

Manage hot and cold memory by loading fast-access hot context by default and retrieving cold summarizations on demand. Use spawn templates or flags like --include-cold to optionally load historical data when current context becomes bloated during multi-session conversations.

Agents load full context history on every spawn because they lack a tiered memory model to separate active data from durable knowledge. Implementing distinct hot, cold, and wiki memory tiers enforces selective loading to prevent payload bloat and slow decision-making.

You can load wiki knowledge on-demand during debugging by using the --include-wiki flag or spawn templates. This keeps durable reference knowledge separate from immediate hot context, ensuring the agent retrieves deep documentation only when needed.

Tiered memory works for agent onboarding and knowledge consolidation by preserving durable wiki knowledge while maintaining fast access to current hot context. It improves latency and memory efficiency across multi-session tasks by isolating immediate data from long-term references.