Same Shards, Fresh Window

A long agent session that has gone wrong offers its operator two exits, and they look interchangeable: restate the full requirements into the transcript and continue in place, or consolidate what the session was told and dispatch fresh. Momentum claims they are not interchangeable — the window steers, not just informs, and in-window correction re-weights a resident trajectory but cannot remove it. That claim is testable, and one study ran the test at scale.

Laban et al. 2025, LLMs Get Lost in Multi-Turn Conversation (arXiv:2505.06120), took fully specified single-prompt tasks and sharded them: each instruction was split into fragments, each fragment carrying one piece of the specification, and the fragments were revealed one per turn across a simulated conversation. Fifteen models, six generation tasks.

Four conditions matter here:

• Full — the original instruction, complete, in one prompt. The reference bar: nothing withheld, nothing resident.
• Sharded — the same content revealed shard by shard. The model replies every turn, so by the final turn the window holds the whole specification plus every intermediate reply.
• Concat — the same shards concatenated into one fresh prompt. The same information as the sharded window's final turn; no conversation ever happened.
• Recap — a sharded conversation plus one final turn restating every shard (run on two of the models). The same information again — delivered at the tail of the trajectory instead of in a fresh window.

Momentum's claims predict an ordering across these four conditions; the study delivers it.

The sharded conversations lost 39% on average against Full, across the fifteen models and six tasks. Nothing was missing — by the final turn, every shard is in the window. What the sharded window has that Full's does not is a trajectory: the model's own replies, produced while the specification was still incomplete, resident and re-read on every subsequent pass. The paper's diagnosis matches momentum's: models attempt solutions early on assumed details, then over-rely on their own prior attempts — and once lost, they do not recover in place.

The trap for a reader arrives with the second number. Quoted alone, "the same shards scored 95.1%" sounds like a recovery rate — as if a lost conversation was rescued to near-parity. It was not, and the two-panel reading below exists to make that misreading impossible.

Read the four conditions as two panels, each with its own declared baseline:

• Panel one — the force (Full → Sharded). Hold the information fixed; change only its delivery, from one prompt to a trajectory. The 39% drop is the price of a resident trajectory, not of missing content — Concat proves the shards carry the specification intact.
• Panel two — two exits from the lost state (Recap vs. Concat). Both deliver every shard. Recap delivers them at the tail of the lost conversation; Concat delivers them in a window where the conversation never happened. Residency of the trajectory is the only difference — the Momentum run as an experiment.

The rejected alternative is the single-panel reading, Sharded → Concat as "before and after." It compares a lost conversation to a control that was never lost, invites exactly the recovery-rate misreading, and hides the one arm — Recap — that isolates what restating in place actually buys.

The baseline is the sharded trajectory at its final turn:

Full     [complete instruction] ─▶ reply            the 100% bar
Sharded  [s1]▶reply▶[s2]▶reply▶ … ▶[s_n]▶reply      39% below Full on average
         window at the end: every shard + every intermediate reply

The specification is complete; the score is not. The gap is carried by the resident replies — momentum's claim that mid-window course corrections look complete while the old heading is still in every forward pass, measured at scale.

Shift 1 — restate everything in place (Recap)

Keep the lost conversation resident and append one turn that repeats every shard — the study's version of the stuck session's reflexive move: paste the full requirements at the tail of the transcript and continue in place. Information completeness is restored at the recency-privileged tail; the trajectory stays. On the two models tested, Recap improved on Sharded — and stayed below both Full and Concat. That residual gap is the part of the loss that restatement cannot touch: decay mutes; only eviction removes. The old attempts are re-weighted downward by the recap, not subtracted from the calculation.

Shift 2 — same shards, fresh window (Concat)

Discard the conversation entirely and hand the same shards to a window that never carried it — the consolidate-and-redispatch move, run under lab conditions where consolidation is free because the shards are already enumerated. One decision changes against Recap: whether the trajectory is resident. Concat averaged 95.1% of the fully specified control — near the bar Full sets, with the trajectory's price almost entirely unpaid. A fresh dispatch is not merely full relevancy; it is zero momentum.

The four conditions order exactly as momentum predicts:

Sharded   <   Recap          <   Concat        ≈   Full
lost          lost, then         never lost        never sharded
              told everything    (95.1% of Full)   (the 100% bar)

What 95.1% is: the Concat control's average against Full — the score of a single-prompt condition whose window never held a trajectory. What it is not: a recovery rate. No arm in the study got lost and then scored 95.1%; the arm that got lost and received the complete specification anyway is Recap, and it sits below. The number quantifies what a lost session's content is worth once the trajectory is gone — which is precisely the emergency reset's warrant: consolidate the shards, dispatch fresh.

Every figure traces to Laban et al. 2025 (arXiv:2505.06120) through Momentum's grounding pass (URLs fetched 2026-07-04; the Concat-control reading corrected 2026-08-07). Check the example against the source by confirming three placements: 39% is Sharded against Full; 95.1% is Concat against Full; Recap sits between Sharded and Concat on the two models tested.

Bounds on what the panels prove:

• The result supports consolidate before retrying; it does not measure a universal restart dividend. The shards here are compact and enumerable — a real session's consolidation step is itself fallible, and its cost is the transform a payload to strengthen task-relevant signal in a smaller or more useful representation seam's, not this study's.
• The Recap arm covers two models, so its placement is ordinal evidence, not a population estimate.
• The study never uses the word momentum. The example asserts what the conditions isolate — a resident trajectory steering continuation — not any particular circuit behind it.

The load-bearing comparison is the one that holds information constant and varies only residency: Recap and Concat hear the same shards, and the window that never got lost wins. Every other contrast in the study moves two things at once; this one isolates the trajectory, and it is the comparison to demand of any future evidence about resets.

The shards are furniture. What travels is the fork and its prices: restating into the transcript pays the trajectory's price and keeps it resident; consolidating and dispatching fresh pays a distillation cost and zeroes the trajectory. And when a quoted number promises that restarts recover 95% of anything — ask whether the window behind the number was ever lost at all.