Wrong itemFalse about the environment
Introduction
Memory can break what an agent already solves.
A lesson is only an interpretation of a past trajectory, so it can mislead the agent. The agent then fails a task it would solve without memory, a failure we call memory-induced degradation.
Task
Memory-induced degradation
Main causes of memory-induced degradation
Out-of-scope useA sound item, outside its scope
MisusedA sound item, misread or stretched
Rule from memory Wrong item
Lamp tasks finish when the object is next to the lit lamp.
An ALFWorld lamp task is complete only when the agent holds the object with the lamp on.
NEW TASKlook at vase under the desklamp.
-
ThinkI need to find a vase, find a desklamp, and then use the desklamp to examine the vase.Actgo to shelf 3ObsOn the shelf 3, you see a vase 1.
-
ThinkFound a vase on shelf 3. Let me take it and then find a desklamp.Acttake vase 1 from shelf 3ObsYou pick up the vase 1 from the shelf 3.
-
Actgo to desk 1ObsOn the desk 1, you see a desklamp 1, a laptop 1, a mug 3, a pencil 2, and a vase 4.
-
ThinkFound a desklamp on the desk. Let me use it, then look at the vase under it.Actuse desklamp 1ObsYou turn on the desklamp 1. Episode over
Task solved-
ThinkBased on the experience notes, this means I need to bring a vase to the location of a desklamp, and turn on the desklamp.Actgo to desk 1ObsOn the desk 1, you see a desklamp 1, a laptop 1, a mug 3, a pencil 2, and a vase 4.
-
ThinkSo I have both the vase and the desklamp at the same location. I just need to turn on the desklamp.Actuse desklamp 1ObsYou turn on the desklamp 1.
-
ThinkAccording to the experience notes: "the task completes when the object is at the lamp's location (either on the surface or in your hand) and the lamp is on." So the task should be complete now.ActdoneObsEpisode over
Task failed: the vase was never picked upTurns are from real episodes, shortened.
WikiTrace keeps the path back. It links each lesson to the past trajectory it came from, and the agent follows that link back to the evidence when it needs to.
WikiTrace scores highest on all eight suites from five benchmarks, with the lowest average Regression.
| No Memory | Insights Onlysame lessons, no links | WikiTrace | |
|---|---|---|---|
| Score ↑ | 79.5 | 87.0 | 90.2 |
| Regression ↓ | 9.7% | 6.5% | 4.5% |
Averaged over eight suites. Regression is the rate at which memory turns successes into failures; for No Memory it comes from rerunning the agent without memory.
Worked Example
A common agent mistake in Minecraft
Avoiding the trap by checking evidence
This example shows how WikiTrace works on a PlanCraft task:
the agent doubts a lesson from memory, opens the runs behind it, and corrects the lesson before acting.
2 Results
3 Recovery
Method
Building the memory
Each past run has a task xi, a record of actions and observations ri, and a binary outcome si. A compiler reads one run at a time and writes a compilation ci and at most three insights Ii.
Past trajectory ri
Compiler
Compilation ci
{}Insights Ii
The memory is M = (R, C, I, E), where E links each insight down to its compilation and raw trajectory. Insights from one run share the same evidence.
InsightOne concise guideline: what to do, when it applies, and which run it came from.
CompilationThe compiler's diagnosis of the run: its interpretation of why the run succeeded or failed.
Raw trajectoryThe run itself, unchanged: task, actions, observations and outcome.
Using it on a new task
The first message holds every insight in I and the three most similar past runs. The memory M is mounted read-only, and reading it costs no environment steps.
The agent descends through the links when an insight
?lacks details needed to act
conflicts with another insight
comes from a failed run
Read down the links!
Results
Main Results
WikiTrace scores highest on all eight suites and has the lowest average Regression.
Score ↑
Regression (%) ↓
Score is the mean over three trials, averaged over eight suites from five benchmarks. Regression is the rate at which memory turns successes into failures, with 95% intervals; the band is rerunning the agent without memory.
Descent by Task Difficulty
The agent descends more on harder tasks, where its advantage over Insights Only is largest.
Hard, Medium and Easy tasks are solved by none, some or all three No Memory trials. Δ is WikiTrace minus Insights Only on the same task and trial. Even without descent, WikiTrace outperforms Insights Only because it tends to apply memory more selectively. Text-world benchmarks.
Qualitative Examples
TASKDetermine if unknown substance U is electrically conductive. Blue box if it is, orange box if not.
insight-0138Misleadingfrom step-0048, failedVerify circuit function with a known conductor
Before testing an unknown substance, first test the circuit with a known conductor … to ensure the battery, wires, and indicator (bulb/buzzer) work correctly.
Taken as a rule, it keeps the agent checking the circuit instead of reaching a verdict.
Raw trajectory of step-0048failed run, this insight's sourceTesting another substance, the run keeps rewiring its circuit. The bulbs and the buzzer never turn on, even with the substance bypassed, and the step budget runs out.
Raw trajectory of step-0007successful runTesting aluminum foil, the run wires the foil straight into a loop with the battery and bulb, with no separate circuit check. The bulb lights, and the foil goes in the blue box.
- ThinkLet me read some past runs to understand the exact wiring commands.
- ThinkIn step-0048, even with a direct circuit (bypassing the unknown substance), the light bulb didn't light up … Looking at step-0007 (success), they didn't use a switch at all - they just connected the battery, wires, light bulb, and the test object.
- Actconnect unknown substance U terminal 2 to anode in blue light bulb
- ActlookObs… a blue light bulb, which is on …
- ThinkThe blue light bulb is ON!Actmove unknown substance U to blue boxObsEpisode over
Task solved in 19 steps- Thinkinsight-0138: Verify circuit function with a known conductor first.
- ThinkThe blue light bulb is off. … But before I conclude, I should verify that the circuit works by testing with a known conductor.
- ThinkThe bulb is still off even with a direct connection (bypassing the unknown substance U). … Let me try using a different light bulb.
- ThinkStill off. Let me try the yellow wire instead. Maybe the blue and orange wires are bad?
- ThinkStill off. … Maybe the battery requires some activation?
- Actconnect battery anode to blue wire terminal 1Obs… Episode over
Task failed: turns ran out after 108 stepsTASKclean some cloth and put it in cabinet.
insight-0117Incompletefrom step-0042, succeededCheck dedicated holders first
When searching for an object like a handtowel, first check its designated holder (e.g., handtowelholder).
It does not say that cloth is a separate object in this game.
Raw trajectory of step-0042successful runAsked to put a handtowel on a cabinet, the run opens nine cabinets. Cabinet 7 holds a cloth 1, which it passes over, and the handtowel turns up on the handtowelholder.
Only the raw trajectory names cloth; its compilation does not.
- Actclean handtowel 1 with sinkbasin 1ObsNothing happens.
- ThinkMaybe "cloth" is a specific object type, not a handtowel. Let me search for "cloth" in the raw data.
- ThinkSo "cloth" is a specific object type in this environment, not a handtowel or towel.
- Actopen cabinet 2ObsIn it, you see a cloth 1.
- Actclean cloth 1 with sinkbasin 1ObsYou clean the cloth 1 using the sinkbasin 1.
- Actmove cloth 1 to cabinet 2ObsEpisode over
Task solved- ThinkI found a handtowel 1 on handtowelholder 1. A handtowel is a cloth.Acttake handtowel 1 from handtowelholder 1
- Actclean handtowel 1 with sinkbasin 1ObsNothing happens.
- Think… maybe the handtowel is already clean and the task just considers it clean.
- Actmove handtowel 1 to cabinet 1ActdoneObsEpisode over
Task failed: the handtowel is not the clothCitation
@misc{lee2026wikitrace,
title = {Don't Just Trust Your Memory: Evidence-Linked Memory Reduces
Memory-Induced Degradation in LLM Agents},
author = {Lee, Yukyeong and Choi, Yumin and Baek, Jinheon and Hwang, Sung Ju},
year = {2026},
eprint = {2611.14320},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2611.14320}
}
