When the harness decides success: state that cannot be re-observed

A support or data assistant repeats work across related customer requests.

What you would change

Try the paper's memory change in a replay of past tasks, with separate memory per customer.

Try the paper's memory change in a replay of past tasks, with separate memory per customer.

Compare against

Your current memory pipeline plus a no-memory control using the same agent.

1. Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.
2. Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.
3. Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.
4. Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test.

What to measure

Task success, stale memories, retrieval misses, response latency, and all memory maintenance costs.

When to keep it

Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation.

recommendation

test

research area

memory and context

technical change

Context-window summarization with note-taking for agents

published

2026-10-19

The mechanism is cheap to implement on constrained infrastructure and the pre-registered synthetic result identifies a concrete failure mode, but the production-relevant magnitude rests on single-run estimates and one model on four games, so workload-specific validation is required before adoption.

1. claim 01Paperraft inference
   Read the brief interpretation

   The method replaces drop-oldest-turn context truncation in agent harnesses with summarization of prior turns plus an explicit invitation for the model to take notes. It costs additional tokens per turn for summary generation and added harness complexity, and provider-native reasoning replay may be needed to recover the remaining gap. It can fail when a required fact is never re-observable and the summary or notes omit it, and the ARC-AGI-3 numbers are single-run point estimates that may not transfer to other benchmarks or models.

   This is a Paperraft inference. It is not pinned to a verified PDF passage.

   Inspect the source
2. claim 02Paperraft inference

   In a pre-registered synthetic probe, truncation costs nothing when facts can be re-observed but drops success to chance when they cannot (penalties differ by 92.5 percentage points, 90% CI [89.5, 95.5]); on ARC-AGI-3, summarizing old turns instead of dropping them recovers 86% of a 63-point harness gap, though from single counted runs per configuration.

   This is a Paperraft inference. It is not pinned to a verified PDF passage.

   Inspect the source