A clean history can mislead the next agent
TL;DR. Salesforce AI Research (Egor Pakhomov, Shafiq Joty, Erik Nijkamp and colleagues) freezes tool-using agents mid-task with their exact world state saved. They then edit either the history or the world and let a receiver model continue. The only difference between arms is the edit. The same correct content can help or hurt depending on how the history states it. On AgentCE-Bench, after the agent commits a wrong value, adding the right value as a truthful note beside the preserved error raises recovery from 51% to 66%. Adding the same value as if the agent had completed it lowers recovery from 50% to 44%, worse than no edit. On AppWorld, a "success" receipt for a database write cuts whole-task success from 45% to 28% when the write's effect is absent, and raises it from 38% to 49% when it is present.
flowchart LR
E["Agent error<br/><small>wrong value committed</small>"] --> H["History edit<br/><small>compaction or repair</small>"]
H -->|note beside error| N["Honest history<br/><small>error stays visible</small>"]
H -->|rewrite as done| C["Clean history<br/><small>error hidden</small>"]
W["World state<br/><small>still holds the error</small>"] --> N
W --> C
N --> G["Recovery 66%<br/><small>from 51%</small>"]
C --> B["Recovery 44%<br/><small>from 50%</small>"]
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
class E,W input
class H loop
class N,C core
class G exit
class B err
linkStyle 5 stroke:#2f9e44,stroke-width:2px
linkStyle 6 stroke:#e03131,stroke-width:2px
Key findings
- Form beats content. A truthful note next to the preserved error: 51% to 66% recovery. The same value phrased as the agent's own completed action: 50% to 44%.
- Receipts must match effects. A success receipt with no actual effect: 45% to 28%. With the effect present: 38% to 49%. The 27-point interaction is positive in six of eight receiver models at 97.5% confidence.
- Practical rule. Test compaction, repair and handoff policies against the real world state.
How it relates to prior wiki pages
- Directly qualifies Folio (11-15, which folds agent history into coarser recorded views and ran 2.4x cheaper on LOCA-bench). Folio's views are recorded by the runtime, so they describe what happened. This paper shows what goes wrong if a fold or summary describes a cleaner world than the real one.
- Confirms the memory-tax result from 10-28 (memory tax, where rewriting history dropped facts and scored below raw-history retrieval) and WikiTrace (11-12, which links each stored lesson back to its raw trajectory). Three papers now say the same thing: keep the evidence, annotate it, do not overwrite it.
- Counterfactual method. Freeze, edit one thing, continue. This is the same intervention design as MemCredit (11-15, which values each memory by whether removing it changes the next action) and CoReMem (today).
Gaps
- Two benchmarks and a handful of receivers. Long-horizon tasks with many errors, where compaction pressure is highest, are not tested.
- No proposal for an automatic consistency check between history and world. The rule is stated, not implemented.
Links
- Paper: https://arxiv.org/abs/2611.14654
- Raw:
raw/dair/2026-11-16-123502-dair-academy.md - Concepts: agent-memory · agent-harness-engineering