Skip to content
Rishabh Salian
  • Work

    Flagships · 2025–26

    • Follow
    • Healthy Materials
    • Housing Works
    • Greener Hours

    Archive · 2019–23

    • Stun Gun
    • Lotus Heater
    • YAAP
    • OBC Bank
    • VSG
    • Music Rooms
    • BEST Bus

    Early art · 2016

    • Analogies boards
    • Charcoal works
  • About
  • Contact
  1. Home/
  2. Work/
  3. Follow
the invisible teammate

Capstone · MS Strategic Design & Management, Parsons · Spring 2026

Follow.

One shared memory for your team’s AI chats, and the artifacts they produce.

Follow sits between the AI tools your team already uses and turns every chat and document into shared knowledge: one index of what the team knows, and a directory of who knows it.

Skip to the live demo↓

Project timeline

Spring 2025 → May 2026: field research, expert interviews, and the two pivots that turned Follow from an AI-native document tool into a team-memory layer.

Field researchExpert interviewProduct momentPivot
PIVOT 1 · MAR 17doc editor → workspace ext.PIVOT 2 · LATE APRindividuals → teams-firstSPRING ’25FEB ’26MAR ’26APR ’26MAY ’26Housing Works3-person team · the originConcept v1cognitive debt · AI-native docPeer tests4 peers · early wireframeAlexandra Beckerconcept test · outside view · Mar 31WorkshopIntelligence Gap · 4 teamsAndyexpert · agentic systemsBuild sprintsMCP tools · the pipelineAtanu Sinhaexpert · 25-yr operatorv6 capstoneteams-first · this artifact

What this is not: a longitudinal deployment study. The peer tests informed a pivot, not a validation; the longitudinal pilot ahead is the move from modeled to measured.

The research

Mixed-methods, and honest about scope: a theoretical spine, plus six primary engagements, from lived experience to expert interviews.

Wegner · transactive memory systems
Clark & Chalmers · the extended mind
Sweller · cognitive load theory
Bienefeld et al. · TMS in human-AI teams
Kosmyna et al. (MIT) · cognitive debt
Doshi & Hauser · homogenization risk
Housing Works NYCLived experience
Intelligence Gap workshop4 teams · 12 participants
Peer concept tests4 peers · early wireframe
AndyExpert · agentic systems
Alexandra BeckerConcept test · outside perspective
Atanu SinhaExpert · 25-year operator

The problem

As work moves into AI workflows, the reasoning behind it disappears into private chats.

Every decision now gets worked out with an AI first: the constraints, the rejected options, the why. All of it stays in one person’s thread, invisible to the rest of the team.

Teams already run as a transactive memory system with each other: everyone keeps a rough map of who knows what. Nothing like it exists for their AI tools: the context that shaped the work scatters across separate chats. Across five concept tests, most people volunteered the same feeling unprompted: their AI-assisted work didn’t quite feel like theirs.

The structural change · today

Same team. Same tools. The reasoning scatters.

MMayaClaudeAAlexChatGPTSSamGemini
Team docslive · uploaded
✕no shared memory
the only place it meetsThe meeting.partial recaps · no sources · forgets what it doesn’t know
Each teammate works with their own AI in a private thread, the team’s docs sit off to the side, and the only place it all meets is a meeting, which forgets what it doesn’t know.

Adapted from the capstone deck: three teammates, three AIs, each working in a private thread the others can’t see.

Insights & areas of opportunity

Three findings shaped where Follow plays, and where the opportunity is largest.

01

AI is the invisible teammate.

Teams already run as transactive memory systems with each other, but not with their AI, because its contributions were never captured in a form the team could route to.

02

Cross-tool memory is structurally vacant.

Native memory inside one AI tool is solved by vendors. Cross-tool, cross-contributor memory is empty; no vendor with the surface to build it has a reason to make it cross-vendor.

03

Provenance matters where stakes are high.

The strongest signal for value comes where the cost of being wrong is asymmetric: legal review, regulated work, compliance. That’s also where pricing tolerance is highest.

“The deliverable shipped — we didn’t fail. But I could have done better work if I’d known how my teammates got where they got.”

From the Housing Works experience; the five concept tests kept surfacing the same feeling, unprompted.

The opportunity

How might we give a team one shared, trustable memory, across every AI tool they already use?

The response

Follow.

A shared memory layer that lives between your AI tools, not inside any one of them. This is the whole system:

MASMaya · ClaudeAlex · ChatGPTSam · Geminiguest checkout · threadfee modelling · threadSCA & retries · threadusability-notes.pdfanalytics-export.csvpayment-sheet spec · livecheckout · topic41% drop at shipping revealSCA: no retry codeLCP −0.8sAlex’s LTV readWCAG blockerstrust signalsspec fact · v1guest default · Mayaspec fact · v2
“default to guest checkout” lands as a decision from Maya’s threadtyped edges: produced by its thread · tagged to the checkout topicit contradicts Alex’s LTV read: red tie, both stay livethe spec doc re-syncs: v2 supersedes v1; the old version stays in the traila query pulls it back out, answer with author + source attachedone graph: chats and documents, every version keptthe team’s memory as a graph: chats and documents in, typed edges, versions kept
Maya · Alex · Samfactconversationfile · uploaded · livetopicproduced · authoredtopic tagcontradicts→supersedessuperseded · kept

the whole system

Three teammates, three different AI tools, and the team’s documents, with one shared memory between them. Hover or tap any part to see what it does, and click the index to look inside it.

What it does

Not another search box over your chats. A memory with receipts, and six things fall out of that.

typical RAG

Finds text that matches.

You get back a similar-sounding paragraph: no owner, no date, no idea whether the team still believes it.

Follow

Finds the source.

Who worked it out, in which chat, when, what it connects to, and whether anyone disagrees.

vs

Glean indexes your documents. Follow indexes the reasoning that produced them.

Maya’s Claude, Alex’s ChatGPT, Sam’s Gemini: every teammate’s AI reads and writes the same memory, over MCP.

Every fact carries who said it, in which chat, and when, so answers come back attributed, and you can check them.

When two teammates’ AIs conclude different things, Follow flags the conflict and keeps both sides on the record.

The directory of what your team knows and who knows it, maintained by the work itself, not by anyone filling in profiles.

New decisions retire the old ones they replace, and the old version stays in the trail, so nothing silently vanishes.

Uploads and live docs alike. Follow follows the artifact: its facts land in the same shared memory and version up as it changes.

The sandbox

The shipped dashboard, replicated live.

the story

01

The team

Aurora is a fictional checkout-redesign team. Maya designs in Claude, Alex runs product in ChatGPT, Sam builds in Gemini: three tools, normally three silos.

02

The week

One working week, captured as they worked: 16 conversations, 7 files, a 32-fact memory. Three questions ended the week still contested; Follow keeps both sides on the record.

03

Your seat

You’re the fourth teammate. Ask Follow anything about the week: it thinks, picks its tools, and you watch every call cross the wire. Save the conversation and you join the memory too.

new here? the tour: five stops, two minutes

live model · Follow's actual MCP tool contracts (Follow ↗) · fictional sample workspace

↻ app.follow.team/w/aurora-checkout
Follow.sandbox
Aurora — checkout redesign3 teammates · 3 AI tools · one memory
memory synced · 32 entries

views

the team

MMayaClaude11
AAlexChatGPT11
SSamGemini10

this workspace

Facts captured32
Topics tracked8
Contested pairs3
Last capturetoday

All items

Every captured conversation, uploaded file, and extracted fact: one feed, newest first.

loading the workspace…

MCP console

Follow is headless by design: the shipped server exposes these tools over JSON-RPC at /mcp, to people and machines alike. Same names, schemas, and response shapes here.

A real model, running Follow's real tools. It decides which to call; every call and result crosses the wire below; the answer is grounded in what came back.

⚙ query_index⚙ directory_query⚙ detect_contradictions⚙ get_activity⚙ save_conversation⚙ retrievesource: Follow ↗

You've got the fourth seat on Aurora; Maya, Alex, and Sam's week is already in the memory. Try one of the prompts below: the model thinks, picks its tools, and you'll see the JSON-RPC-shaped traffic in the open. Ask it to save the conversation and this session appears in Conversations and Facts, under “You”.

Ask Followlive model
Ask about Aurora's week: what's contested, who to talk to, what changed. I check the team's index before answering, and you'll see every step.
Concept sandbox · pre-loaded sample workspace · the shipped Follow product captures real AI threads over MCPanswers run on a live model API via a server-side proxy

Views

the team

MMayaClaude11
AAlexChatGPT11
SSamGemini10

this workspace

Facts32
Topics8
Contested3

open the sandbox full-screen ↗

The sandbox above runs the shipped system, not a mockup. What’s behind it:

Stack: a TypeScript monorepo (Turborepo, pnpm workspaces): a Hono MCP server exposing 12 tools over JSON-RPC at /mcp, a Next.js dashboard, two browser extensions (activity capture, Google Workspace), a folder-watching desktop agent, and an Expo mobile shell. Deployed on Railway.
Pipeline: five LLM roles (Reporter, Analyst, Editor, Archivist, Profiler; the Editor currently flag-gated) turn raw AI conversations into a versioned index: typed edges, provenance on every fact, supersession instead of overwrites.
Testing: Vitest, ≈1,350 automated tests across 165 files as of July 2026, covering the pipeline, tool contracts, and clients.
Ownership: the architecture, data model, tool contracts, and evaluation are mine; implementation is AI-assisted with Claude Code, with every change reviewed and tested.
Limits: beta scope; capture quality tracks MCP client support.
← PreviousGreener HoursNext →Healthy Materials

Rishabh Salian · portfolio · 2026

  • Email
  • LinkedIn
  • GitHub