Feature Proposal: Tiered History Retrieval
Problem
Agent history.md files grow unbounded. Currently Scribe summarizes at 12KB, but agents read everything — no distinction between recent context and archived knowledge. This wastes context window tokens and slows agent startup.
Proposed Solution
Formalize a two-tier retrieval pattern inspired by ClawMongo's tiered projection model:
Hot layer (history.md): Last ~20 entries. Always read at spawn time. Fresh, actionable context.
Cold layer (history-archive.md): Summarized older entries. Read only when agent explicitly needs historical context.
Implementation
- Spawn prompt change: agents read history.md (hot) by default. Only read history-archive.md when the task references past decisions or old patterns.
- Scribe summarization: when history.md exceeds threshold, move old entries to archive with summaries (already partially implemented).
- Search cold: add a 'search archive' pattern agents can invoke when they need historical context.
Benefits
- Reduced context window usage (~60% less tokens for history)
- Faster agent startup
- No external DB required — pure file-based
Prior Art
Inspired by ClawMongo's tiered retrieval model (https://clawmongo.site) which uses ID-only projections for 10x context reduction. Adapted for Squad's file-based architecture.
Feature Proposal: Tiered History Retrieval
Problem
Agent history.md files grow unbounded. Currently Scribe summarizes at 12KB, but agents read everything — no distinction between recent context and archived knowledge. This wastes context window tokens and slows agent startup.
Proposed Solution
Formalize a two-tier retrieval pattern inspired by ClawMongo's tiered projection model:
Hot layer (history.md): Last ~20 entries. Always read at spawn time. Fresh, actionable context.
Cold layer (history-archive.md): Summarized older entries. Read only when agent explicitly needs historical context.
Implementation
Benefits
Prior Art
Inspired by ClawMongo's tiered retrieval model (https://clawmongo.site) which uses ID-only projections for 10x context reduction. Adapted for Squad's file-based architecture.