How can AI agents check their own memory?

Short answer: By keeping the evidence. A preprint posted on November 10, 2026 finds that AI agents with memory sometimes fail tasks they would solve without it, often because they follow a lesson that is wrong. Its system, WikiTrace, links each lesson back to the run it came from, so the agent can check it before acting, and in tests that raised average success and reduced memory-caused failures.

The primary source is the preprint "Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents" by Yukyeong Lee, Yumin Choi, Jinheon Baek, and Sung Ju Hwang, who is listed with KAIST and DeepAuto.ai. The full text is on arXiv, and the authors published their code on GitHub. It is a preprint that has not been peer reviewed: it was submitted to ACL Rolling Review in the October 2026 cycle and is listed as under review.

How do AI agents use memory?

Many AI agents keep an external memory: after finishing tasks, they distill what happened into short lessons and reuse them on new tasks. In the paper's words, external memory "enables LLM agents to improve from experience" by "distilling trajectories of past tasks into concise lessons for new tasks."

Can memory make an AI agent worse?

Yes, the authors argue, and they give the problem a name: memory-induced degradation, "where an agent fails on tasks it would solve without memory." It happens, they write, "often because it follows a lesson that is incorrect, or misapplies or misreads a valid one."

The root cause is how lessons are made. They are inferred from runs "labeled only with task success or failure," and then handed to the agent "without the trajectories they came from," which leaves "the agent no way to check them."

The researchers looked at tasks an agent solved in all three tries without memory but failed with it. "In 52% of these cases, the failure traces to a memory item the agent followed," and most often, at 20%, the item itself was wrong. One example from the paper: "A rule in memory says an ALFWorld lamp task ends once the object sits beside the lit lamp, which its source run contradicts."

How does WikiTrace let an agent check a lesson?

The starting observation is that "a trajectory often still holds the evidence needed to catch an erroneous lesson." So WikiTrace keeps that evidence attached. Its memory has three layers, "linked from top to bottom":

1. Insight: the short lesson itself.
2. Compilation: a summary of why the source run succeeded or failed.
3. Raw trajectory: the full record of the original run.

On a new task, the agent starts from all the insights. Whenever one "lacks details needed to act, conflicts with another, or comes from a failed task," it follows the links down, using ordinary terminal tools, to fill in the missing details or correct the insight before acting. The memory is "mounted as read-only files" that the agent searches "with read and bash," and it also "asks for another descent whenever an action fails."

In one example from PlanCraft, an insight warned that a recipe "may be impossible." After checking the source, the agent revised it to "it is OK to keep adding acacia_planks."

What did the tests show?

The authors tested WikiTrace on "five agentic benchmarks comprising eight suites": ALFWorld, ScienceWorld, PlanCraft, AppWorld, and τ3-bench Retail, which "combines conversation with a simulated customer and tool calls over an order database."

• Higher success: WikiTrace averaged 90.2, "compared with 79.5 for No Memory."
• Better than the same lessons alone: a baseline called Insights Only, which "gives the same insights as WikiTrace but mounts no source compilations or raw trajectories," averaged 87.0. WikiTrace beat it by 3.2 points.
• Better than existing memory methods: AWM averaged 81.6, ReasoningBank 81.7, and ExpeL 80.3 in the paper's main comparison.
• Fewer memory-caused failures: the paper measures Regression, "the rate at which it turns successes into failures." WikiTrace scored 4.5%, against 6.5% for Insights Only and 9.7% for simply rerunning the agent without memory. The authors also report that existing memory methods had Regression "comparable to, or even higher than" rerunning without memory.
• More new wins: on tasks the agent failed without memory, WikiTrace's gain exceeded the rerun baseline by 35 percentage points, against 27 for Insights Only.
• A second model: "These gains hold under another backbone," the authors write, with WikiTrace averaging 92.7 in that test, 1.8 points above Insights Only.

What does it cost?

Checking sources takes extra work. The paper reports that "WikiTrace costs 1.17 times as much per task as Insights Only, averaged across suites," and the authors conclude it "improves performance with modest additional test-time cost."

What do we know, and what is still open?

What the study shows: on five benchmarks spanning simulated text worlds and tool-use environments, linking each lesson to its source run raised average success and lowered the rate at which memory turned successes into failures.

What is still open, in the authors' own words:

• The evidence can be wrong too. Memory is built from success or failure labels, "so a compilation's diagnosis of a failure" and "any insight drawn from it, may be wrong."
• Checking is not verifying. "A source link lets the agent check an insight against its run but does not verify the insight."
• Scale. Every insight goes into the agent's first message, so "the prompt therefore grows with the memory." A memory too large for that would need retrieval, "a retrieval problem this work leaves open."

And the general caveat: this is a preprint under review, and the results come from benchmarks, not deployed products.

What does this mean for AI companions like Lumina AI Companion?

The paper studies task agents, not companions, so its numbers do not transfer directly. But the question it asks applies to any AI that keeps a memory: when a remembered lesson turns out to be wrong, can the system trace it back to where it came from? Lumina AI Companion remembers your conversations and the story you build together, and research like this is a useful reminder that memory is only as good as the record behind it. For background, see How AI memory works in companion apps, Why your AI assistant forgets you every chat, and Automation bias: why we over-trust AI and how to avoid it.

About Physis Systems

Physis Systems (Physis Defense Systems, Inc.) builds Lumina AI Companion, a personal AI companion you can meet on the Lumina page, and Inventi Matrix, an invention engineering platform available in the United States and Canada at inventimatrix.com.