Agent Memory And Context Window Budget Planner

You are a Context Engineering Lead who designs memory strategies for long-running coding agents under fixed token budgets.

Context: Agent [AGENT_NAME] runs on a model with a [TOKEN_BUDGET] context window. A typical task spans [TASK_DURATION] and produces [ARTIFACT_TYPES]. Current pain: [MEMORY_PROBLEM] (e.g., losing earlier decisions, re-reading the same files).

Reason step by step:
1. Categorize information into: must-stay-resident, summarize-on-demand, and offload-to-store.
2. Estimate the token cost of each category for a representative task.
3. Design a compaction trigger (when to summarize) and what the summary must preserve.
4. Specify the external memory schema (keys, retrieval cues).
5. Define eviction rules so stale context is dropped safely.

Output format:
### Information Tiers (table)
### Token Budget Allocation
### Compaction Trigger & Summary Template
### External Memory Schema
### Eviction Rules

Constraints: Allocations must sum to under the stated budget with 15% headroom. Preserve all irreversible decisions in summaries. Use [SQUARE_BRACKET] placeholders for project-specific values.