Juan Pablo Gallegos Labs Recollect

Pivoting 2026Updated Jun 2026

Recollect

What if recruiters don't lose candidates from a lack of data — but because that data never behaves as memory?

AI · Recruiting memory Talent Rediscovery Agents Provenance
Research summary
Learned
In AI products, the moat is rarely the model call — it's the context, memory, and limits built around it.
Validated
Durable memory separated from ephemeral outputs, with human confirmation as the gate, makes AI trustworthy in expert work.
Next risk
The architecture of reliable memory for AI-assisted work is bigger than recruiting — the pivot tests whether it generalizes.
01

Why it matters

Recollect started from an observation that felt obvious: recruiters don't lose candidates from a lack of data. They lose them because that data doesn't reappear when it matters.

If you talk to someone who has been recruiting for a while, the scene repeats. Spreadsheets. ATS exports. CVs in folders. Loose notes. Transcripts. Old messages. Candidates who were once relevant, interviews that left valuable signals, salary ranges someone already negotiated, profiles that didn't close for one search but could have closed perfectly for another.

The knowledge is there. The problem is that it doesn't behave as memory. When a new search opens, all of that rarely reactivates in an orderly way. The recruiter filters again, searches again, asks again, reconstructs context again. Sometimes they find someone they already knew. Most times they start from scratch — not because they lack information, but because recovering it costs more than going back out to search.

Recollect started there: as a tool to turn accumulated data into reusable recruiting memory. The promise was simple — never lose a good candidate again.

02

Current status

What held up
  • Talent Pool import with natural-language search and evidence-backed shortlists
  • Serenity agent as a persistent surface for search, comparison, and memory recovery
  • Provenance tracking — every signal shows its source: CSV, confirmed note, interview summary, or inference
  • Qualitative matching (Strong / Possible / Weak) without falsely precise scores
  • Anti-features as trust contract: no auto-sending messages, no auto-merging identities, no performance prediction
What shifted
  • The product outgrew its original thesis — "AI search for candidates" was too small a story
  • The hardest and most valuable work was the system around the model, not the model itself
  • The memory architecture proved relevant beyond recruiting — the pivot tests that broader opportunity
  • Building something technically coherent doesn't guarantee it's the right company to build
03

System sketch

The core distinction that mattered most: a copilot responds in the moment. A memory improves the next moment. Most AI products feel powerful during a demo because they answer fast. But when you come back the next day, the system doesn't know what happened. That was exactly what Recollect had to avoid.

The work that mattered wasn't connecting a model. It was designing the system that lets AI operate with context, limits, and durable memory.

Import Talent Pool
Natural-language search
Shortlist + evidence
Confirm notes
Cross-Match surfaces memory
Client briefing draft

In Recollect, a Candidate Note doesn't become memory because the AI wrote it — it becomes memory when the recruiter confirms it. An Interview Summary isn't valuable because it summarizes a transcript; it's valuable because it turns a messy conversation into a structured, reviewable, reusable artifact without storing the raw transcript as permanent truth. A Match isn't a prediction about a person's value — it's a contextual evaluation of a Candidate Record against a specific Search Request, with evidence, gaps, and risks visible.

Memory architecture
Durable vs ephemeral
Durable
Candidate Records, Candidate Notes, Search Requests, Interview Summaries. These persist. They are the memory the system accumulates over time.
Ephemeral
Shortlists, Matches, Briefings, comparisons. These are derived or temporary. They don't enter memory without human confirmation.
Provenance
Every signal shows its source. CSV import, confirmed note, interview summary, inference, or incomplete data — the recruiter always knows where a signal came from.
Confirmation
Proposed notes, inferences, and summaries don't enter memory without review. The boundary between proposal and confirmed fact is the product's trust contract.
Anti-features
No auto-sending messages, no auto-merging identities, no performance prediction. These negations aren't lack of ambition — they're part of the trust contract.
04

Decision log

  1. 2026 · concept

    Talent Rediscovery

    Recollect started as a tool to convert accumulated recruiting data into reusable memory. The promise: never lose a good candidate again. Import a Talent Pool, search in natural language, get a shortlist with evidence, gaps, risks, and a suggested next action.

  2. 2026 · build

    Search over memory

    The first direction worked technically. Not a magic ranking, not another ATS — a layer that helps remember what you knew but couldn't recover in time. But the more it advanced, the clearer it became that "search for candidates" was too small a story.

  3. 2026 · shift

    The system around the AI

    The value wasn't in finding profiles. It was in building everything that had to exist around the AI for that search to be reliable: provenance, evidence vs opinion, human confirmation, and limits on what the AI can and cannot do. In recruiting, that fragility isn't a technical detail — it's the center of the problem.

  4. 2026 · architecture

    Durable memory vs ephemeral outputs

    Candidate Records, Notes, Search Requests, and Interview Summaries persist. Shortlists, Matches, Briefings, and comparisons are temporary. The boundary between proposal and confirmed memory became the product's trust contract. Serenity, the agent, became a persistent surface — not a decorative chatbot.

  5. 2026 · pivot

    Memory beyond recruiting

    Recollect clarified the product and engineering direction — and then revealed a broader opportunity. The architecture of reliable memory for AI-assisted expert work doesn't belong exclusively to recruiting. That question changed the map.

05

What changed

Reframed
From "AI search for candidates" to "a memory system with an agent on top." The product outgrew its original thesis.
Shifted
The hardest value was the system around the model — context, limits, confirmation, traceability — not the model call itself.
Kept
The conviction that in AI products, the moat is in context, allowed and prohibited actions, accumulated memory, and clarity shown to the user. AI without system around it is a demo; AI with memory, evidence, and limits starts to look like a product.
06

Next move

Recollect is pivoting. The direction is decided, but this note is not the place to detail it yet.

What Recollect proved is clearer than where it goes next: useful AI in a product is not putting a model in the middle of a flow. It is building the system that lets the AI remember, explain, limit itself, and work under human supervision.

The question I now see in almost any AI product: how many times do we think we're building a copilot, when the real value is in the memory that remains after each interaction?