mediar-ai Context Compression Strategies
Compress agent context while preserving critical information for long-running sessions.
Prerequisites
- Understanding of context windows
- Familiarity with summarization techniques
Instructions
Key Insight
Optimize for tokens-per-task, not tokens-per-request. When compression loses critical details, agents must re-fetch information, wasting more tokens overall.
Compression Approaches
| Approach | Compression | Quality | Trade-off |
|---|---|---|---|
| Anchored Iterative | 98.6% | 3.70 | Best quality |
| Regenerative | 98.7% | 3.44 | Moderate |
| Opaque | 99.3% | 3.35 | Best compression |
Structured Summary Sections
## Session Intent
[What the user is trying to accomplish]
## Files Modified
- file.ts: Description of changes
## Decisions Made
- Key decision and rationale
## Current State
- Progress status
## Next Steps
1. Next action item
Compression Triggers
| Strategy | Trigger Point |
|---|---|
| Fixed threshold | 70-80% context utilization |
| Sliding window | Keep last N turns + summary |
| Task-boundary | Compress at logical task completions |
Probe-Based Evaluation
Test compression quality by asking questions:
| Probe Type | Tests |
|---|---|
| Recall | "What was the original error?" |
| Artifact | "Which files have we modified?" |
| Continuation | "What should we do next?" |
Guidelines
- Use structured summaries with explicit sections
- Implement incremental merging, not full regeneration
- Track artifact trail separately (weakest dimension)
- Accept slightly lower compression for better quality
- Monitor re-fetching frequency as quality signal
Notes
- Structure forces preservation through dedicated sections
- Artifact trail integrity is universally weak (2.2-2.5/5.0)
- Sliding window with structured summaries is best for coding agents
Source: muratcankoylan/Agent-Skills-for-Context-Engineering