---
name: progressive-cramming
description: Evaluate whether compressed AI-agent context preserves decisions, constraints, entities, and unresolved work across successive summaries. Use for context-window management, memory compression, and long-running agent workflows.
---

# Progressive Cramming

Compress context in stages while protecting the information that future work depends on.

## Workflow

1. Identify non-negotiable constraints, decisions, named entities, open questions, and current state.
2. Compare the original context with its compressed version.
3. Mark each critical item as preserved, weakened, contradicted, or missing.
4. Calculate retention by weighting constraints and decisions above background detail.
5. Rewrite the summary to restore missing high-value information without restoring low-value history.
6. Repeat the check after every major compression cycle.

## Output

Return retention score, lost critical items, contradictions, and a repaired compact context.

Prefer explicit state and decisions over narrative chronology.

## Research basis

[Progressive Cramming: Reliable Token Compression and What It Reveals](https://arxiv.org/abs/2607.21231) — Tarasov et al., arXiv (2026). The paper introduces the exact term for learned token compression; this skill provides a simplified retention diagnostic rather than reproducing the paper’s method.