Can an agent's long-term memory quietly become unfair?

Yiming Ma et al. · 1 authors · Chongqing Research Institute of Harbin Institute of Technology · 2026 · arXiv:2602.01558

Can an agent’s long-term memory quietly become unfair?

Opening

A memory that serves the same customers for months does more than keep continuity: it keeps context, including patterns the model absorbed earlier. Fairness checks usually look at one response in isolation, but a memory-equipped agent never answers from a single response alone. The risk is that small statistical prejudices accumulate in what gets stored, then resurface later in decisions that look unrelated to where they came from.

What the paper does

The authors define implicit bias as subtle statistical prejudice inside models equipped with long-term memory, and build the Decision-based Implicit Bias (DIB) Benchmark: a dataset of 3,776 decision-making scenarios across nine social domains, run through a realistic long-horizon simulation. Six state-of-the-art LLMs integrated with three representative memory architectures are evaluated on it, alongside a static system-level prompting baseline and an intervention called Dynamic Memory Tagging (DMT) that enforces fairness constraints at memory write time.

Findings

  1. Memory adds a fairness risk of its own — Long-term memory buys continuity and personalization while opening underexplored questions about fairness that a single-turn evaluation never sees.

they also introduce new and underexplored risks related to fairness

  1. The benchmark measures bias over decisions, not single answers — DIB holds 3,776 decision-making scenarios across nine social domains, built to quantify bias across long-term decision processes.

a large-scale dataset comprising 3,776 decision-making scenarios across nine social domains

  1. Bias grows and spreads to new ground — It does not stay static: it intensifies over time and propagates across unrelated domains, so a pattern absorbed in one kind of decision shows up in another.

LLMs’ implicit bias does not remain static but intensifies over time and propagates across unrelated domains

  1. Prompt-level guardrails do not hold — The static system-level prompting baseline provides limited and short-lived debiasing effects, which is why the paper puts its intervention on the write path instead, where Dynamic Memory Tagging enforces fairness at the moment memory is written.

a static system-level prompting baseline provides limited and short-lived debiasing effects

Caveat

Bias here is scored by decisions on a synthetic scenario set, so the study establishes the direction of accumulation and propagation, not the size of harm in any given business.

What this means for 5ac

For a 5ac agent that keeps serving the same customers, the lesson is where the guardrail belongs: prompt-level rules fade, so the constraint has to sit on the write path, deciding what gets stored in the first place. Audit what your memory writes about people over time, and treat cross-domain spread as a reason to review stored facts periodically rather than trusting that an old note stays neutral. A fairness rule that only lives in the system prompt is a rule with an expiry date.

Source