Faded Memory (Dynamic Context Management)

Re-render older observations at lower detail with deterministic, local decay policies. Allow agents to “refresh” details on demand from ground-truth state.

Concept

• Apply exponential/linear decay to token budgets per observation as turns grow.
• Re-render from Versioned state snapshots; no LLM summarization needed.
• Allow active refresh (e.g., printing a variable) to bring details forward.

Example Policy (exponential)

• Most recent: full budget
• Previous: 1/2 budget
• Earlier: 1/5 budget

Benefits

• Long-running conversations with graceful degradation.
• High fidelity by re-rendering from source state.
• Cheap and deterministic vs. LLM summaries.

Considerations

• Integrate with system token counting and visibility.
• UX for refresh hooks; caching for performance.

Related issue: Issue #5