# Hiring Assistant 2 Learns Your Past Hires. That Can Cut a Pool 27x

_LinkedIn Hiring Assistant 2 learns from your past hires and passes. Here's how its memory can shrink a PM pool 27x, and what to audit before it switches on._

- Canonical URL: https://www.refolk.ai/blog/hiring-assistant-2-memory-audit-past-hires
- Published: 2026-11-01
- Reading time: 10 min
- Keywords: LinkedIn Hiring Assistant 2, Hiring Assistant 2 memory, LinkedIn Recruiter AI agent review, LinkedIn Hiring Assistant verification signals, AI sourcing bias past hires

LinkedIn Hiring Assistant 2 is about to treat your past hires as your default search, and you won't see it happen in the search box. It switches on automatically this month for English-language customers, at no added cost, and its headline feature is a memory of how you and your company have hired before. If your last 20 hires came from one industry, one seniority band and one commute radius, the agent is about to treat that pattern as what you meant.

## What does Hiring Assistant 2 memory actually learn?

Hiring Assistant 2 memory is a persistent record of your hiring behavior, kept per recruiter, per role and per company. The agent uses it to guess what you meant when you type a search. It learns from who you hired, who you invited and who you passed on.

LinkedIn announced the update at Talent Connect in New York. LinkedIn's stated goal was an agent that does "what I meant, not what I said." Dan Reid, LinkedIn's VP of Product, gave the fuller framing, and it's the one to keep in mind:

> It understands more about how you hire, not how everybody hires, but how you hire.

Three details from the coverage matter more than the keynote language:

- **Scope is wider than one person.** HeroHunt describes it as recruiter-level *and* company-level memory. A colleague's habits may shape your results.
- **Inputs include negative signals.** SocialFind reports that the people you invited and the people you passed on both shape what it suggests next.
- **Your ATS can feed it.** Teams that connect an ATS let the agent evaluate their whole talent pool, including people who applied outside LinkedIn. Launch integrations are Greenhouse, Lever, SmartRecruiters, Zoho Recruit and Tracker.

That last point means years of old ATS dispositions, not just this quarter's LinkedIn clicks, can become material the agent learns from. And as ALM Corp points out, the announcement does not explain how the memory is stored or how recruiters can edit what the assistant has learned.

## How much can a remembered history shrink a candidate pool?

In Refolk's index of professional profiles, a memory trained on retail product-manager hires starts from about 3.5% of the US product-manager market: 2,374 people out of 67,494. That's roughly a 27x cut, and the recruiter never typed it into the search box.

The example in HR Briefing's coverage of the launch shows how this works. A recruiter searches "product manager," and the agent returns mid-level retail PMs near the office, because that's who the recruiter hired before. I couldn't find that example in LinkedIn's own materials, so treat it as illustrative rather than a quote from LinkedIn. The mechanism it describes matches how LinkedIn describes memory.

Here is what that inference does to the numbers:

| Segment | US profiles | Note |
|---|---|---|
| All "Product Manager" titles | 67,494 | Refolk's index |
| Product managers at Retail-industry companies | 2,374 | Refolk's index |
| Product managers outside Retail | 65,120 | 67,494 minus 2,374 |
| Retail share of all US PMs | 3.5% | Derived |
| Non-retail pool vs. retail pool | ~27x | Derived |

> **3.5%** Share of US product managers who work at Retail-industry companies
>
> In Refolk's index, 2,374 of 67,494 US PM profiles. A retail-trained memory starts here before location narrows it further.

The shape of the pool changes too, not just its size. Retail PMs in the index cluster at a handful of names: Starbucks, Kroger, Lowe's, The Home Depot, Sephora and Nordstrom. Across all US PMs, no employer dominates. The top employers are a scattered list that includes Ramp, Brex, Intuit, Wiz, Formlabs, SKIMS and Sherwin-Williams. A memory that learned from retail hires can keep pointing at the same six logos, and over time your shortlists start to look like copies of your last hires.

Two caveats on the data. Title matching is broad, and the retail sample picked up some off-target titles such as "VP Operations," so read these as close estimates. A query for retail PMs tagged "Senior" returned zero, which is almost certainly a seniority-tagging gap rather than a real absence.

### Why the efficiency gain hides the problem

LinkedIn says recruiters using version 1 reviewed 83% fewer profiles before finding a qualified match. That number is the reason this bias is hard to see. When a shortlist arrives fast and every name on it looks plausible, nobody asks about the 65,120 people the agent decided not to show. A narrower funnel *feels* like a better one.

The fix is to measure the pool somewhere the memory can't reach. Before you trust a remembered default, run the same brief somewhere neutral and compare sizes. [Refolk](/) is built for exactly that check: you describe the person in plain English, it searches without your past hires weighting the results, and you can see what the full market looks like next to what the agent chose to show you.

> **Example search:** "Product managers in the US with 4 to 8 years of experience who have shipped B2B or fintech products, excluding retail companies" - [run it on Refolk](https://www.refolk.ai/s/x2741q0ksg)
>
> A ranked shortlist from a search your past hires don't weight, so you can compare the full pool against what a history-trained agent surfaces.

## Why "who you passed on" teaches the agent as much as who you hired

Passes are signals too. If memory learns from invites and rejections, then every rushed "not a fit" click becomes a preference the agent may carry into future searches.

This is the part most teams will miss, because nobody audits their passes. Think about how dispositions actually get made:

1. **High-volume reqs.** On a req with hundreds of applicants, recruiters bulk-reject on thin evidence. Those clicks reflect time pressure, not judgment.
2. **Requirements that changed later.** A role that once required on-site attendance produced a pile of rejections for remote candidates. If the policy changed, the rejections still say what they said.
3. **Legacy ATS rejection codes.** Once Greenhouse, Lever or SmartRecruiters is connected, old reason codes like "overqualified" or "location" may become patterns the agent learns from.
4. **One recruiter's taste.** Because memory is also company-level, one teammate's strong preference for a single school or employer can spread into everyone's results.

The mechanism is simple. A recommender treats a rejection as evidence about the *type* of person, not about the moment. A pass that meant "not now, the req is full" turns into "not this profile."

## Why memory works against changing what you hire for

Memory is designed to repeat the past, so it works against any team trying to hire differently from how it hired before. HeroHunt puts it bluntly: a company trying to change its talent profile is asking the agent to do the opposite of what memory optimizes for.

This hits hardest for three kinds of teams:

- **Teams entering a new market**, such as a retail company building a fintech product line.
- **Teams changing seniority**, such as moving from mid-level hires to staff and principal roles.
- **Teams running career-switcher or returnship programs**, where by definition the right candidate doesn't look like your past hires.

For these teams, a history-trained default is a slow leak. The req stays open, the shortlists keep showing people you've already decided against, and after a month someone suggests bringing in an agency. The workaround is to restate the change in the brief every time ("not retail," "staff level and above," "career changers from teaching or consulting welcome") and to source the first slate somewhere your history isn't steering the results.

## Audit your last 20 hires first: a pre-switch checklist

Before the update reaches your seat, review your last 20 hires and last 20 passes per role family. Write down what pattern they show, and decide whether you want the agent to repeat it. Twenty is enough to spot a pattern and small enough to finish in an afternoon.

Work through this list:

1. **Pull your last 20 hires per role family.** Note industry, seniority, location and previous employer. If 15 of 20 came from the same industry, assume the agent will treat that as your intent.
2. **Pull your last 20 passes on the same reqs.** Flag any that were bulk rejections, policy-driven, or made under requirements that no longer apply.
3. **Check what your ATS will expose.** If you plan to connect Greenhouse, Lever, SmartRecruiters, Zoho Recruit or Tracker, ask your admin what disposition history comes along with it, and clean up the reason codes that are stale.
4. **Write a "what changed" line for every open req.** One sentence on how this hire should differ from the last one. Paste it into every brief.
5. **Measure the unfiltered pool.** For your two or three most important roles, size the market outside LinkedIn so you have a baseline to compare the agent's shortlists against.
6. **Spot-check shortlists monthly.** Count how many surfaced candidates share an employer or industry with your past hires. If that share keeps rising, the memory is narrowing your pool.

Steps 5 and 6 need a pool count that doesn't come from the tool you're checking. That's the practical case for running a parallel search in Refolk on your hardest reqs: you get a ranked list you can compare against LinkedIn's, and you'll notice when the two start to diverge.

## What does a LinkedIn verification badge actually prove?

A verification badge confirms that a member's identity, and in some cases their employer, is real. It does not confirm skill, and because members can hide badges, a missing one tells you almost nothing.

LinkedIn is pushing verification hard, and the fraud problem behind it is real. CEO Dan Shapero has framed the update partly around trust, and LinkedIn's internal research backs him up:

> **64%** US talent acquisition professionals who say it's getting harder to tell who and what is real
>
> LinkedIn's internal research. LinkedIn says 115 million members have verified their identity so far.

Oscar Rodriguez, LinkedIn's VP of Trust Product, is expanding verification to include colleague confirmation. Social Media Today notes that users can opt out of displaying these badges. HeroHunt adds the key limit: verification confirms identity and employment, not skill.

That creates a quiet second filter. If recruiters, or the agent's ranking, start treating verified profiles as safer bets, people who opted out of the badge slip down the list. That group may lean toward privacy-conscious or senior people, the ones who get enough recruiter messages already and choose not to advertise. Treat a badge as one input, and verify skill where it actually shows up:

- **Engineers:** code they've shipped and projects they maintain in public.
- **Product managers:** shipped launches, public talks, written product teardowns.
- **Everyone:** employment confirmed through references or a second source, not inferred from a badge.

## Write your search criteria down outside LinkedIn

If the memory can't be inspected or edited, your hiring criteria need to live somewhere you control. Otherwise the clearest record of how your team hires will sit inside a vendor's model.

HeroHunt points out the trade-off: memory makes the agent smarter every month you use it, and harder to leave every month too. That isn't a reason to avoid it. It's a reason to keep your own copy of what you know:

- A one-page brief per role family: must-haves, nice-to-haves, explicit exclusions, and the "what changed" line.
- A list of target companies and adjacent industries you actually want to pull from.
- Notes on why your best past hires worked out, written in your own words, so you aren't relying on the pattern the agent inferred.

That document is portable. You can paste it into LinkedIn Recruiter, into an ATS search, or into any plain-English search tool, and get comparable results whichever tool runs it. When the agent's guess about "what you meant" drifts, the written brief is what you correct it against.

## FAQ

### When does LinkedIn Hiring Assistant 2 turn on?

LinkedIn began rolling out Hiring Assistant 2 in November after announcing it at Talent Connect in New York. Customers who use the product in English get the update automatically at no added cost, so most teams won't have to opt in. Because the memory learns from your hiring activity, the audit is worth doing before the update reaches your seat.

### Can I edit or delete what Hiring Assistant 2 memory has learned?

LinkedIn hasn't said publicly. ALM Corp's review of the launch notes that the announcement doesn't explain how the memory is stored or how recruiters can edit it. Until LinkedIn documents this, assume you can't reliably undo it. Clean up your ATS dispositions before you connect the ATS, and restate any change in hiring direction in every brief.

### Does AI sourcing bias from past hires really shrink the pool that much?

It depends on how concentrated your history is. In Refolk's index, retail-industry product managers are 2,374 of 67,494 US PMs, about 3.5%. A memory that infers "retail" from past hires would start roughly 27x smaller than the full market before location narrows it further. Teams with more varied hiring histories will see less narrowing, but the only way to know is to compare the agent's shortlists against an independent count.

### Should I filter Hiring Assistant 2 results to verified profiles?

Not as a hard filter. Verification confirms identity and sometimes employment, not skill, and members can choose to hide their badges. Filtering on badges risks dropping privacy-conscious and senior candidates who opted out. Use the badge as one fraud check among several, and verify skill through work you can actually see, such as shipped code, shipped products and references.

---

_From the Refolk blog. More posts: https://www.refolk.ai/blog/index.md_