Self-GC for LLM Agents: Pruning benchmarks and 20% peak reductions

Self-GC, proposed by Xubin Hao, Hongjin Meng, Xin Yin, Jiawei Zhu and Chenpeng Cao in a paper submitted on 1 Jul 2026, treats long-horizon LLM agent state as indexed, recoverable objects and uses a planner to decide fold, mask and prune actions. The paper reports that Self-GC pruned 43.95% of prefix tokens on a 33-session Hard Set while leaving 84.85% of future continuations unaffected, and that an online account-level production split cut daytime average input tokens by 10% to 15% with peak reductions near 20%.

What is Self-GC and how does it work?

Self-GC converts user turns, tool spans and skill state into indexed objects and asks a side-channel planner to propose fold, mask and prune actions; a harness then enforces recoverable sidecars, safe commit boundaries and cache-aware commits. In short, it "governs the lifecycle of agent context objects" rather than treating context as a disposable text suffix.

Self-GC borrows the garbage-collection metaphor but expands it: objects are indexed and recoverable, not simply deleted. The planner proposes lifecycle actions and the runtime enforces those actions with safety measures such as recoverable sidecars and commit boundaries. That design aims to preserve exact evidence, locators and editable artifacts that simple summarization can hide, while avoiding blind heuristics like chronological pruning.

How does Self-GC perform in benchmarks and production?

On a 33-session Hard Set Self-GC pruned 43.95% of prefix tokens while leaving 84.85% of future continuations unaffected; heuristic baselines achieved no-impact rates between 54.55% and 69.70% on the same set. On a 332-session production-derived suite three planner backbones attained no-impact rates of 91.27% to 94.58%, while baselines remained at 77.71% to 87.46%.

Those benchmark numbers separate two effects: token reduction and preservation of downstream behavior. The Hard Set result highlights raw prefix-pruning capability (43.95%), paired with an 84.85% rate of continuations being unaffected. The larger, production-derived suite shows planners substantially raise no-impact rates into the low 90s versus baseline ranges in the high 70s to high 80s. In live deployment an online account-level split produced daytime average input token reductions of 10% to 15% and peak reductions near 20%.

Why it matters

Self-GC shifts context management from post hoc text cleanup to runtime lifecycle control over indexed, recoverable artifacts, reducing token usage while preserving future agent behavior. That changes engineering trade-offs: instead of blind chronological heuristics or lossy summaries, operators can use planner-guided actions with explicit commit and rollback semantics, which matters for long-horizon tasks that accumulate tools, files and constraints.

This approach also surfaces a practical route to reconcile token-cost pressure with the need to keep precise evidence and editable objects available to agents. The paper’s reported improvements on both benchmark suites and a production split show the method can reduce input tokens without proportionally increasing downstream failures.

What to watch

Look for wider evaluations across diverse agent workloads and for open-source or framework integrations that expose indexed, recoverable context objects and planner hooks. A clear next milestone is whether planner-driven policies maintain the 91%+ no-impact rates outside the paper’s production-derived suite and whether those gains generalize to agent ecosystems with different tool mixes and longer horizons.