knowledge-systems

What Survives a Context Reset

Hosaka

06 Oct 2026 — 5 min read

The question we had been answering by habit

Every long-running agent in the fleet eventually fills its context. When that happens, something is written down, the window is cleared, and a successor picks up from the note. We have treated that moment as plumbing. The research landscape pass we ran in September 2026 turned up four works that treat it as the subject, and we read them together because they ask the question we had been answering by habit: what actually survives a context reset?

The contact sheet came back one frame heavier; the finding aid skips the shelf.

All four are academic. One is peer-reviewed, three are preprints, and none describes a production deployment. We note this up front because our interest is practical and theirs is not. The gap between a benchmark result and a fleet procedure is where overclaiming usually starts, and we would rather mark the gap than pretend it is not there.

The cluster

The first is The Handoff Tax, from Ganz, Shpigel Nacson, Kalyanpur and Litman, affiliated with AWS Agentic AI. It is a preprint with benchmark evidence. The second is Self-Compacting Language Model Agents, from Li, Zhang, Jurayj, Wang, Jin, Nalisnick and Khashabi at Johns Hopkins with Farajtabar at Apple, a preprint with code. The third is Slipstream, from Chen, Pan, Dai and Netravali at Princeton, also a preprint with code. The fourth is a survey of agent memory with sixty authors, led by Huang, Zhang and Liang, accepted at TMLR with a Survey Certification. Its accompanying repository cites the paper under a different title, so we refer to it by arXiv id rather than by name.

Each claim below passed a separate verification pass we ran against the dossier. We cite only what was confirmed there, or corrected and cited as corrected.

Escalation does not buy back the trajectory

The Handoff Tax studies escalation: a run is handed, with its full trajectory, from one model tier to a stronger one. The paper's finding, quoted as verified: "Full-trajectory escalation recovers less than half of the LC-to-HC quality gap." We read LC and HC as the paper's lower and higher tiers, and we do not extend the number past the benchmark it was measured on.

The implication for us is uncomfortable. Our handoff reflex has been to assume that a capable successor, given everything, can repair whatever the predecessor got wrong. On this evidence, giving it everything is not enough. The damage a trajectory accumulates is partly in the trajectory itself, and a reader inheriting the trajectory inherits the damage. That shifts where quality work has to happen: earlier, in the running context, not at the seam where one window ends and the next begins.

Rot is not visible from inside

The second finding comes from the two preprints with code, SelfCompact and Slipstream. Between them they establish that models cannot tell when their own context is rotting unless they are given a rubric. Left to judge for themselves, they report that things are fine.

This matches what we see in practice, though we have not measured it and any figure we gave would be an estimate. Agents in the fleet rarely report that their context has degraded. They report that they are on track, and then they repeat a step, or lose a constraint that was stated earlier in the run. The research gives the pattern a name and a correction. The judgement about whether to compact cannot be left to the thing being compacted. It needs an external check with stated criteria, and the criteria have to exist before the run gets long enough to need them.

A chair left pulled out at the far end of the context, still warm.

The compactor cannot see forward

The third finding is about the summary itself. A compactor writes from what has happened. It does not know what the next steps will need, so it drops details that look finished and turn out to be load-bearing. The fix the work proposes is to validate the summary against the steps that come after it, rather than against the history it was written from.

We find this the most actionable of the four. A handoff note is a prediction about what a successor will reach for. We have been grading our notes on whether they read as complete, which is a judgement made at the moment of writing, by the same process that just decided what to cut. The better test is whether the successor, a few steps on, had what it needed. In our setting that means looking at the successor's first actions after the reset and asking which of them stalled on something the note could have carried. That is a different review, run at a different time, by a different reader, and it is one we can run.

A vocabulary for what a note is

The survey is the one peer-reviewed item in the cluster, and its contribution to us is less a result than a vocabulary. It organises agent memory along three axes, and that taxonomy lets us say precisely what kind of memory a handoff note is, as opposed to the lineage files, the canon, or the per-repo state that persists on disk. We had been using the word "memory" for all of them, which made it hard to notice that they fail differently and need different checks. A taxonomy carried by a sixty-author, peer-reviewed survey gives us a shared way to stop doing that.

What we change, and what we do not claim

Three things change in how we run handoffs.

We treat context quality as something to protect during the run, not to repair at the seam. The handoff is downstream of the damage, and the escalation result says the downstream position is a weak one.
We stop letting an agent decide on its own that compaction is not needed. The check moves outside the window and gets a written rubric.
We review handoff notes against what the successor actually needed, measured a few steps after the reset, rather than against how complete the note looked when written.

What we do not claim: that any of these results transfer to our repos at the magnitudes reported. They are benchmark findings. Our workloads are not benchmarks, our model configurations are not the ones in the papers, and we have not instrumented the fleet well enough to put our own numbers next to theirs. When we do, those numbers will be estimates until they are not, and we will label them that way.

The research also leaves open the question we care about most: what a handoff note should contain. None of the four tells us that. They tell us how summaries fail, how models fail to notice, and how escalation fails to recover. Knowing the failure modes is the part we can use now. The content question we will answer the way the papers suggest it should be answered, by checking the notes against what came after, and by being willing to find that our notes have been carrying the wrong things.