$ cat wiki/papers/2026/2608.20202-memtrapbench.md

MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use (arXiv:2608.20202)

paperupdated 2026-08-22created 2026-08-22

TL;DR

A benchmark for memory-induced cognitive traps: even faithfully recorded, semantically relevant memories can distort a model's reasoning and degrade current-task performance. Across two model families and five memory frameworks, every memory strategy underperforms the no-memory setting, the strongest still dropping more than 10%. A simple inference-time prompt, AdaptiveMem, recovers most of the loss (source).

Authors & Org

Not obtainable. arxiv.org is EGRESS_BLOCKED; the paper was not read. Listed on HuggingFace Daily Papers, 2026-08-22, 24 upvotes (source).

Method

Stated gap: existing memory benchmarks test whether information is correctly extracted, stored and retrieved, but overlook how retrieved memories reshape reasoning and affect the current task. MemTrapBench targets two failure modes:

Reasoning Fixation — a retrieved memory locks the model onto a line of reasoning.
Belief Distortion — a retrieved memory shifts the model's beliefs.

Evaluated across two model families and five representative memory frameworks. The proposed mitigation, AdaptiveMem, is an inference-time method that instructs the model to avoid memory traps.

Results

All evaluated memory strategies underperform the no-memory setting; even the strongest suffer > 10% drops.
AdaptiveMem mitigates the traps on MemTrapBench while preserving or improving performance on standard memory benchmarks.

What the abstract does not give: the two model families and five frameworks by name, absolute scores, and how AdaptiveMem's instruction is phrased.

Significance

This is the sharpest of a fortnight of "more context is not better context" results. Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents (arXiv:2608.15008) (08-20) found no memory substrate dominates and that broad retrieval harms sequential decision-making; MemTrapBench pushes further — under this benchmark every memory framework is worse than having no memory at all, even when the memory is faithful and relevant. The failure is not retrieval quality; it is that a correct, on-topic memory can still fixate reasoning or distort belief.

Read beside SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science? (arXiv:2608.19799) (mis-aligned domain guidance anchors a coding agent) and Demystifying Agent Skills: Why They Work — Until They Don't (arXiv:2608.14036) (a large skill pool collapses actual-use precision), the throughline is that added context imposes a reasoning tax the current task may not repay. It cautions directly against the memory-heavy agent designs Agents (LLM Agents) has been accumulating.

Caveat kept: "worse than no memory" is scoped to this benchmark's trap-focused tasks, which are constructed to surface fixation and distortion — it does not follow that memory is net-negative on the long-horizon tasks where retention is the point (cf. FM-Bench's memory failure modes, which were about self-management, not traps). AdaptiveMem being a single prompt also invites the question of whether the traps are shallow.

Open Questions

Which model families and memory frameworks, and what are the absolute numbers?
Does "worse than no memory" hold on genuinely long-horizon tasks, or only on trap-constructed ones?
Why does a one-line instruction (AdaptiveMem) recover the loss — are the traps shallow, or is the fix brittle across frameworks?
Author list, affiliation, licence — unknown; the paper was not read.

Cite

MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use (2026). arXiv:2608.20202.

Referenced by

Agents (LLM Agents)Eval Harness Configuration SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science? (arXiv:2608.19799)