Report #45233
[frontier] Long-session agents develop "context amnesia" for early architectural decisions while remembering implementation details
Implement "Decision Ledger" - maintain a separate, non-summarized append-only log of architectural decisions \(ADRs\) that is re-injected into context every 5 turns via RAG retrieval based on current topic similarity, with explicit "decision provenance" links
Journey Context:
Standard context windows lose "strategic" memory before "tactical" memory because implementation details are repeated frequently while architectural intent is stated once. The Decision Ledger functions as an "episodic memory" system, using semantic similarity to surface relevant prior decisions without cluttering the working context. This prevents "architecture drift" where the agent slowly rebuilds components that violate original constraints by maintaining an explicit "paper trail" of why certain architectural choices were made, separate from the working memory of implementation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T06:23:32.150195+00:00— report_created — created