At a Glance
Link each memory lesson to its original run so agents can check the evidence before following advice; this reduces mistakes caused by bad memories while improving overall performance.
ON THIS PAGE
What They Found
Memory meant to help agents can actively hurt them when a concise lesson hides the wrong cause of past success or failure. linked memory Keeping every lesson linked to its source run lets the agent inspect and revise guidance on demand, turning bad advice into a recoverable mistake rather than a silent failure. Across five benchmarks, linked memory raised average task performance and cut the rate of tasks that memory made worse.
Need expert guidance?We can help implement this
Key Data
1Average performance rose by 10.7 percentage points compared to an agent with no memory.
2Gained 3.2 points over a baseline that had the same concise lessons but no links to source runs.
3Memory-caused failures (Regression) dropped from 6.5% to 4.5%; agents descended to source runs in 22% of Easy and 44% of Hard tasks.
Implications
Engineers building AI agents should use evidence-linked memory to avoid introducing new failures when reusing past runs. Technical leaders and evaluators should track both gains and regressions (what memory fixes and what it breaks) rather than average accuracy alone.
Key Figures

Fig 1: Figure 1: Memory-induced degradation and WikiTrace. (a) The agent solves a task without memory but fails with memory that holds a wrong insight. (b) With WikiTrace, the agent follows the wrong insight’s links to its source run, revises it, and solves the task.

Fig 2: Figure 2: Overview of WikiTrace. (a) A compiler turns each past run into a compilation, a diagnosis of why the run succeeded or failed, and insights, each a concise lesson. All layers stay linked. (b) Each layer’s documents follow a fixed template. (c) The agent starts from every insight and descends through the links under the conditions shown.

Fig 9: Figure 11: Qualitative example: revising a wrong insight (PlanCraft). Insight-0083 leads agents into the auto-craft mistake. Insights Only follows it and fails, but WikiTrace revises it through descent and succeeds.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Memories are built from past runs labeled only success/failure, so the linked diagnosis can still be wrong—the link only makes verification possible, not automatic. Putting every lesson in the agent’s first message grows prompts with memory size, so very large memories need a retrieval step not solved here. Some labels and cause attributions come from a language-model judge and reflect what the agent's reasoning states, not a ground-truth causal analysis.
Methodology & More
Agents that reuse past runs often distill long trajectories into short lessons. Those short lessons can misattribute which action caused a success or failure, so following them blindly causes “memory-induced degradation”: the agent fails on tasks it would otherwise solve. Instead of throwing away concise lessons, [link each lesson to a compact diagnosis and to the original raw trajectory]. The agent starts from every lesson and, when details are missing or two lessons conflict, it “descends” to the linked compilation or raw run to recover evidence and revise the lesson before acting. Researchers built WikiTrace to test this idea. They compile a fixed corpus of past runs (task, full trajectory, binary outcome) into three linked layers: a short insight, a compilation explaining why the run succeeded or failed, and the raw trajectory. On new tasks the agent may consult these links. Evaluated over eight suites across five benchmarks, WikiTrace improved average accuracy by 10.7 points over no memory, beat a baseline given the same concise lessons by 3.2 points, and reduced regression from 6.5% to 4.5%. The agent checks source runs more on harder tasks (22% descent on easy vs 44% on hard), and descents yield larger gains. The takeaway: keep evidence with lessons so agents can verify and adapt guidance instead of trusting memory blindly. short insight raw trajectory
Need expert guidance?We can help implement this
Credibility Assessment:
Contains an established author (h-index 24) and other experienced researchers; although an arXiv preprint, author reputation supports a strong credibility rating.