Ctxpact tackles 100k-token agent prompts locally

Hot take: this looks less like a “summarization” project and more like an execution layer for context triage, and the benchmark story is strongest when it admits that backbone model quality dominates everything else.

• –The 3-stage design is sensible: structural pruning first, summary eviction second, retrieval/extraction last. That ordering reduces wasted LLM calls and preserves the highest-value recent turns.
• –The standout claim is not the compression ratio; it’s the faithfulness result. If Qwen3.5 consistently outperforms LFM2 because it follows retrieved context instead of overriding it with parametric knowledge, that is the real engineering insight.
• –The methodology still needs tighter framing. Frankenstein looks like a narrow, potentially overfit suite, so the 8/8 and 0% variance numbers are persuasive but not yet broad evidence of general robustness.
• –The “2 LLM calls is the sweet spot” result is plausible, but I would want ablations controlling for prompt quality, retrieval candidates, and question difficulty before treating it as a universal rule.
• –LoCoMo-MC10 is a better sign of cross-session usefulness than a single reading-comprehension benchmark, but mixing those scores into a combined percentage can obscure the very different failure modes.
• –The most interesting next compaction ideas are probably hybrid, not deeper agent loops: query-aware hierarchical chunking, structured-field preservation for JSON/tool output, and per-task retrieval policies that choose between pruning, summarizing, and exact recall.