Hiring Assistant 2's Memory Could Hide 39% of Store Managers. Audit Now

If you use LinkedIn's Hiring Assistant, your past invites and passes are about to become the default settings for your next search. Hiring Assistant 2 arrives in November as a free update that needs no migration, and it brings persistent memory. LinkedIn hasn't published the controls for that memory yet. Whatever your history says about who you like, the agent will treat as policy, so check that history before the update lands.

What does Hiring Assistant 2 actually remember?

Hiring Assistant 2 remembers who you invited, who you passed on, and who you hired. It keeps that memory at three levels: per role, per recruiter, and per company. It uses that memory to fill in whatever your search request leaves out.

LinkedIn announced it at Talent Connect in New York on September 29, 2026. Dan Reid, LinkedIn's VP of Product, summed up the change in one line: "before, Hiring Assistant did a search based on what you said. Now, it searches based on what you mean."

In practice, "what you mean" turns out to be "what you did before." LinkedIn's own example is a recruiter who asks for a product manager. The agent returns candidates who live near the office, come from retail, and sit at mid-level, because that's who the recruiter has hired in the past. You never typed "retail." The memory added it for you.

Here's what's in scope:

• Role-level memory: who you've invited and passed on for a given role.
• Recruiter-level memory: your personal pattern of invites and passes across projects.
• Company-level memory: the hiring history of your whole organization.
• ATS data: candidates in your connected ATS (Greenhouse, Lever, SmartRecruiters, Zoho Recruit, Tracker), with iCIMS and Workday coming.

The memory isn't new. Its persistence is.

Hiring Assistant has used memory since the first version. Version 2 widens its scope and makes it persist. That means your 2024 to 2026 project history may already be the training signal. The version 1 charter customers (AMD, Canva and Siemens) have the longest histories.

How much does an "industry like past hires" preference narrow the market?

Even for a role as tied to retail as store manager, an industry anchor hides about 2 in 5 of the market. In Refolk's index, 39.3% of US store managers work outside the Retail industry, and a memory that leans toward "retail, like past hires" would quietly drop all of them.

I picked store manager on purpose, because it's about as retail as a job title gets. If an industry anchor leaks this much here, it will leak more on roles that cross industries.

Source: Refolk's index of professional profiles. I ran the same title search with and without an industry constraint. Title matching is loose, so read these as directional market sizes, not exact headcounts.

74,437

Look at who's in the sample. The top employers include 7-Eleven, O'Reilly Auto Parts and Guess?, but the City of Seattle and a state alcohol commission show up too. Convenience, food service, auto parts and public sector operators all employ store managers, and most of them aren't filed under "Retail." An industry-anchored memory drops that whole tail.

Now apply the same logic to LinkedIn's own example, where the memory carried a retail preference onto a product manager search. Product managers work in every industry, so the share sitting outside retail is almost certainly much larger than 39%. I'm inferring that, not measuring it. But if the narrowing is this big in the easy case, it won't be smaller in the hard one.

Why does a pass teach the agent as much as an invite?

Hiring Assistant 2 learns from who you passed on, not only from who you invited. Every "not now" you clicked tells it something about who you don't want, and the agent can't tell whether you meant "wrong fit" or "wrong week."

That's how the narrowing happens. An agent that learns from who you invited gets better at finding people who resemble who you invited. Once it also learns from your passes, it gets better at hiding people who resemble the ones you skipped. Both pull toward your past.

A typical pass log contains:

• A remote candidate you skipped because the req was briefly marked onsite.
• A career switcher you passed on because the hiring manager wanted a safe first hire that quarter.
• A senior person you skipped because the budget hadn't been approved yet.
• Someone outside your usual industry you never got to because the req closed early.

None of those was a judgment about fit, but any of them can turn into a lasting preference. Because the memory also works at company level, one recruiter's timing-driven passes can shape what a colleague sees on a different req. Every pass you make between now and the November rollout goes into that log.

A pass made for timing becomes a preference made permanent. The agent never learns that it was just a bad week.

The danger is that you won't notice, because a shortlist that looks like your last ten hires looks right to you.

Does the 4x contact rate prove the shortlists got better?

No. LinkedIn says recruiters are 4x more likely to contact candidates sourced by Hiring Assistant than candidates found by traditional methods. That shows the agent agrees with the recruiter. It doesn't show the agent found better people.

If the agent's picks look like people you already chose, you'll reach out to more of them. A memory-driven agent would push that number up even if its shortlists were missing a big slice of the market. Treat LinkedIn's version 1 claim the same way: recruiters review 83% fewer profiles before finding a qualified match. Reviewing fewer profiles is efficient when the filter is right. When the filter is your history, it's just a smaller search.

Speed claims have the same blind spot. LinkedIn says the assistant finishes prescreening in a median of six minutes, against 3.5 hours for a recruiter's first response. But a fast shortlist from a narrowed pool is still a narrowed pool, and none of these metrics measure what was filtered out before you saw it.

20,000+

What should you audit before the November rollout?

Audit the history the memory will read: your project pipelines, your pass reasons, and your connected ATS data. Then write down, in plain language, what you actually want for each live req. The update needs no migration or changes from customers, which also means nobody will prompt you to do this.

Here's the order I'd work in:

1. Pull your project history back to 2024. List the reqs where your hires cluster in one location, one industry or one seniority band. The memory will repeat those patterns.
2. Flag passes made for timing, not fit. Budget freezes, req holds, onsite-only quarters. If the tool lets you archive or relabel those projects before November, do it. If it doesn't, document them so you know which reqs to watch.
3. Separate the reqs where you want to change direction. New office, new market, a role moving from mid to senior. For these you have to tell the agent outright to ignore the patterns it was built to follow, every time.
4. Check what your ATS is feeding in. If Greenhouse or Lever holds years of rejections coded with vague reasons, the memory may treat them as signal. Clean up the disposition codes you can.
5. Write your criteria down outside the tool. Keep a short brief per req: must-haves, nice-to-haves, and the things you deliberately don't filter on. This becomes your baseline.
6. Get answers from your LinkedIn account rep in writing. The questions are listed below.

How do you calibrate an AI sourcing agent against a baseline?

Run the same req twice: once through Hiring Assistant 2, and once through a search with no memory, built only from your written criteria. Compare the size and makeup of the two results. The gap is a direct measure of how much your history is narrowing the shortlist.

Without a neutral reference, you're grading the agent against your own instincts, and the memory learned from those same instincts. So the baseline has to sit outside LinkedIn's memory, or it isn't a baseline. I use Refolk for this. You describe the person in plain English and get back a ranked list that no record of your past passes has shaped. That's how I got the store manager numbers above: one query with the industry constraint, one without.

What to compare once you have both lists:

• Headcount: if the baseline returns far more qualified people, the memory is cutting.
• Industry mix: does the agent's list skew harder toward your past hires' industry than the market does?
• Location spread: LinkedIn's own example favored candidates near the office. Check whether remote or relocatable people disappear.
• Seniority band: does a "senior" req come back mostly mid-level because that's who you hired last year?

Run this check on your first few reqs after November and on every req where you're changing direction. If the gaps stay small, trust the agent more. If they're big, you have evidence to argue with the agent, or with your own past choices.

Who carries the risk if the memory skews your shortlists?

The employer does. LinkedIn provides the agent, but under NYC Local Law 144 and Illinois HB 3773, disparate-impact liability sits with the company making the employment decision. A memory that reproduces past hiring patterns is exactly what those laws are written to catch.

These rules are already in force or on the calendar:

Location, industry and seniority aren't protected characteristics on their own, but they can track ones that are. A preference for "near the office" can mirror neighborhood demographics, and a preference for mid-level can lean on age. If you can't explain what the memory learned, you can't audit it, and these laws expect you to.

Questions to put to your LinkedIn account rep

Public coverage notes that LinkedIn's announcement doesn't explain how the memory is stored or how recruiters can edit what the assistant learned. Ask these before November and get the answers in writing:

• Which Hiring Assistant 2 capabilities turn on by default in November, and which need admin or recruiter opt-in?
• What exactly does memory store per recruiter versus per company, and how do you purge or export it for audits?
• Can individual passes or whole projects be excluded from the memory?
• Does ATS rejection data feed company-level memory, and can that be switched off per integration?
• What audit logs exist to show why a candidate was surfaced or left out?

If the answers are vague, that tells you something too. Lean harder on your baseline searches until the controls are documented.

FAQ

When does LinkedIn Hiring Assistant 2 switch on, and does it cost extra?

LinkedIn is rolling it out in November to existing customers who use Hiring Assistant in English. It's free and needs no migration or changes from customers, so the memory will start shaping your searches without any setup step. LinkedIn announced it at Talent Connect on September 29, 2026.

Can I turn off Hiring Assistant 2 memory?

LinkedIn hasn't publicly documented the memory controls yet. Coverage of the launch notes that the announcement doesn't explain how memory is stored or how recruiters can edit it. Ask your account rep which features are on by default, whether memory can be purged or exported, and whether specific projects can be excluded.

Does Hiring Assistant 2 read my ATS data?

Yes. Candidates in a connected ATS are in scope: Greenhouse, Lever, SmartRecruiters, Zoho Recruit and Tracker, with iCIMS and Workday coming. The announcement doesn't explain how that data is stored or weighted, so clean up vague rejection codes before November and ask your rep whether ATS data can be excluded per integration.

Is AI recruiting agent bias a legal risk for my company?

It can be. NYC Local Law 144 has required bias audits of automated hiring tools since 2023, and Illinois HB 3773 has banned discriminatory AI use in employment decisions since January 1, 2026. Liability sits with the employer, not the vendor, so if the memory reproduces skewed past hiring, the exposure is your company's.