Context Compression

Strategies for long-running sessions exceeding context windows.

Core Insight

Optimize tokens-per-task (total to completion), not tokens-per-request. Aggressive compression causing re-fetching costs more than better retention.

Compression Methods

Method	Compression	Quality	Best For
Anchored Iterative	98.6%	3.70/5	Best balance
Regenerative Full	98.7%	3.44/5	Readability
Opaque	99.3%	3.35/5	Max compression

Anchored Iterative Summary Template

## Session Intent
Original goal: [preserved]

## Files Modified
- file.py: Changes made

## Decisions Made
- Key decisions with rationale

## Current State
Progress summary

## Next Steps
1. Next action items

On compression: Merge new content into existing sections, don't regenerate.

Compression Triggers

Strategy	Trigger	Use Case
Fixed threshold	70-80% utilization	General purpose
Sliding window	Keep last N turns + summary	Conversations
Task-boundary	At logical completion	Multi-step workflows

Artifact Trail Problem

Weakest dimension (2.2-2.5/5.0). Coding agents need explicit tracking of:

Files created/modified/read
Function/variable names, error messages

Solution: Dedicated artifact section in summary.

Probe-Based Evaluation

Probe Type	Tests	Example
Recall	Factual retention	"What was the error?"
Artifact	File tracking	"Which files modified?"
Continuation	Task planning	"What next?"
Decision	Reasoning chains	"Why chose X?"

Six Evaluation Dimensions

Accuracy - Technical correctness
Context Awareness - Conversation state
Artifact Trail - File tracking (universally weak)
Completeness - Coverage depth
Continuity - Work continuation
Instruction Following - Constraints

Guidelines

Use anchored iterative for best quality/compression
Maintain explicit artifact tracking section
Trigger compression at 70% utilization
Merge into sections, don't regenerate
Evaluate with probes, not lexical metrics

2.3 KiB Raw Blame History