Jev Compaction Draws Pushback as Summarization-Style Compression Wins Favor
After mitsuhiko proposed using Jev for on-the-fly context compaction in agents, multiple framework and model authors publicly disputed the approach, sparking a debate over context management strategies that currently favors summarization-based compression.
Confirmed
- Theo rebutted Jev's line-by-line compaction point by point: Jev scores each tool call and drops irrelevant content in place of summarization, but compaction is not a filter.
- pvncher argued, echoing Theo, that Jev cannot access the reasoning trace behind tool calls, so it cannot understand why a tool was invoked; a homegrown compaction scheme also struggles against model-native training strategies.
- Teknium (Nous Research co-founder, Hermes author) said he repeatedly rejected related PRs for Hermes Agent; he explained the scheme deletes 250K tool-call tokens when hitting 500K context, and estimated Jev cannot even cut 50% of context—versus Hermes summarizing 95%.
- When asked to update Hermes for Jev or build a mini harness routing prompts to the cheapest model, Teknium replied Jev cannot help that much with compression.
Why it matters
- The dispute exposes a fundamental split in agent context management: scoring-based pruning versus summarization. Consensus points to pruning without understanding tool-call motivation risking critical context loss, while summarization offers far higher compression rates—a directly useful lesson for long-horizon agent builders.
2026-09-18 ~ 2026-09-20 · 5 related posts
- Episode 1: Jev Compaction Draws Pushback as Summarization-Style Compression Wins Favor(2026-09-18, 5 posts)
- Episode 2: Teknium Tests Jev Compression Scheme, Finds It Worse and Costlier(2026-09-20, 2 posts)
Primary sources
- Teknium explains why "tool-call pruning" compaction keeps getting rejected from Hermes Agent — Teknium ·
- Why swapping in Jev for context compaction underperforms: models are trained on their own strategy — pvncher ·
- Teknium: Jev can't compact context well — Hermes summarizes 95% of it away — Teknium ·
- Theo tears apart Jev's 'instant compaction': line-by-line context trimming misunderstands the fundamentals — AccBalanced · 2026-09-18
- [source] Why swapping in Jev for context compaction underperforms: models are trained on their own strategy — pvncher · 2026-09-18
- Teknium doubts Jev can deliver meaningful context compaction or cheap-model routing — Teknium · 2026-09-19
- [source] Teknium: Jev can't compact context well — Hermes summarizes 95% of it away — Teknium · 2026-09-20
- [source] Teknium explains why "tool-call pruning" compaction keeps getting rejected from Hermes Agent — Teknium · 2026-09-20