BOARDBOX vs. THE CHATBOX

Your AI doesn’t have
a memory problem.
The chatbox does.

A chat window is a brilliant interface for one question and a terrible one for a life. We rebuilt the substrate underneath — channels, profiles, indexes, persistent media, versioned updates — so the model has somewhere to keep what matters.

Back to productSee the difference ↓
The memory problem

What “memory” actually looks like.

Frontier chat apps have improved, but the user-visible memory still often looks like one bucket: saved facts that mix domains, lose scope, and forget to forget. Here’s the shape of a real one — versus a Boardbox profile, on the same person.

Frontier chat app

Saved memories

A single inbox for facts the model decided were important. Weak scope, weak structure, weak expiry. The user sees 1 of 240.

Search memories
…
workWants help wiring their RAG context retriever to Pinecone and building a benchmark harness to test retrieval quality vs. topK and threshold values.
workIs no longer using Temporal for their agentic architecture and will instead use Bull (with Redis) and cron jobs for orchestration.
triviaWears Viktor & Rolf Spicebomb Extreme but finds its longevity lacking.
workIs using tRPC and wants API handlers for messages implemented as tRPC procedures in src/procedures/messages/.
lifeRefinanced the Durham mortgage to 2.9% in 2021; escrow runs about $2,180/mo.
workIs considering using BullMQ's Flow feature to model agent workflows.
healthUlcerative colitis, in remission. Humira biosimilar preferred over the originator.
lifeCurrently earns $8,000 USD/mo and is considering leaving the day job to work on their company.
workWants their agentic workflow to support both text and audio (STT/TTS), with consistent behavior across both.
triviaPrefers iced black coffee in the morning. No sugar.
workIs building an internal AI platform with role-specific agents — CPO, CTO, CMO, CFO — each with scoped tools.
healthCalprotectin 48 on the April panel; next colonoscopy scheduled in October.
Showing 12 of 240 memoriesSearch · Sort · Delete one at a time
Boardbox

Health profile

One of several long-form, structured profiles. The health agent reads this by default — and can pull in your financial profile when you’re pricing a procedure, or your spiritual profile when stress shows up in your sleep. Loaded by relevance, not by topic silo.

HxHealth profileupdated 5/02/2026 · v17 · scoped to #health
HealthPersonalFaithFitness+ New
Primary condition

Ulcerative colitis, in remission — Humira biosimilar, biweekly, self-administered Sunday nights.

Recent labs (Apr 2026)

Calprotectin 48, down from 180 at diagnosis. Next colonoscopy October.

Conditions

OCD (managed). Annual eye exam due Q3. No chronic prescriptions.

5 profiles · diffable · time-stamped · roll back any version
240

Unsorted facts the model has to choose from — every single turn. Work, health, cologne preferences, mortgage details, all flat.

0

Scopes, namespaces, or expiries. The fact that you switched from Temporal to Bull will still surface in a conversation about your calprotectin.

5

Narrative profiles in Boardbox, each domain-scoped. Versioned, diffable, and read-only by the channels that should see them.

The taxonomy

Memory isn’t one problem. It’s seven.

Frontier-chat memory is still constrained by the chat product around it. It fails in seven distinct ways — each one a different shape of the same root cause. We rebuilt the substrate to fix each one by design, not by a bigger saved-memory list.

Architecturally second-class

Memory is a feature. It isn’t the substrate.

ChatGPT, Claude, and Gemini were built around the chat window. Where memory exists, it is layered onto that chat shape. The primary object is still the conversation; memory is an enrichment. In Boardbox the workspace itself — channels, profiles, indexes, artifacts — is the substrate. Conversations are guests.

What it looks like

A flat list of "saved memories" the model scans every turn.

What Boardbox does

A structured workspace the agent walks into already informed.

Cold-start tax

The conversation doesn’t start until turn 12.

Even with memory on, you spend 8–12 turns re-feeding the model context before it’s actually useful. The memory system gives the model fewer questions to ask, not none. Boardbox frontloads your personal profile, workspace index, channel index, and recent thread summaries on turn 1. The model walks in with the room already lit.

What it looks like

Turns 1–12: re-explain who you are, what project this is, what you’re trying to do.

What Boardbox does

Turn-12 quality on turn 1, because the agent already knows.

Scope collapse

A decision inside one project becomes a fact about you.

Tell ChatGPT “I’m using tRPC for this project” and memory promotes it to “you use tRPC.” The next conversation about an unrelated stack assumes tRPC. Project-scoped decisions get globally promoted to identity. In Boardbox, the project, artifact, profile section, and channel are part of the fact, so context can cross-load when relevant without pretending a local decision is universal.

What it looks like

“Wants help wiring their RAG context retriever to Pinecone and building a benchmark harness to test retrieval quality vs. topK and threshold values.”

A decision specific to one project. Now treated as “uses Pinecone” in every conversation.

What Boardbox does

Scoped to the project or artifact that produced it, while still discoverable when genuinely relevant.

Staleness

Facts don’t expire. Ever.

Memory has no model of “is this still true?” Things you said two years ago surface today as current. Career states you’ve already moved on from still show up as “considering.” The system never reconciles.

What it looks like

“Currently earns $8,000 USD/mo and is considering leaving the day job to work on their company.”

True two years ago. The user has been at a different salary for nearly a year. Still surfaces as current.

What Boardbox does

Profiles are versioned and user-editable. When life moves, you update the profile and the agent sees the new state.

Contradictions coexist

Durham AND Brooklyn AND Cincinnati, at once.

Each fact was added during a separate period. Nothing displaced the previous one. Memory holds the contradictions at once. The system has no model of where you are now— only where you’ve been.

What it looks like

“Lives in Durham.” / “Moved from Brooklyn in 2021.” / “Parents are in Cincinnati.”

All three live in memory. No timestamps, no precedence, no reconciliation.

What Boardbox does

Profiles have a single canonical current state. The history lives in versions.

Broken entity continuity

Rename a thing and the model sees two things.

Rename a product, a person, a project, and the memory of the old name doesn’t migrate to the new name. The model treats them as separate entities. You spend energy on every turn forcing the connection. Boardbox makes profiles and artifacts first-class entities with stable IDs — rename the display label, the identity stays.

What it looks like

“User is working on a side project called OtterHalf.” (Renamed to Raft six months ago. Memory still says OtterHalf.)

What Boardbox does

Stable IDs under the surface. Rename is a label change, not an entity change.

Flatness

One bucket for ten domains of your life.

Work, health, finance, faith, dating, fitness, side projects, habits, training — all in one flat memory list, scanned on every turn. Flatness is the reason scope-collapse, staleness, and contradictions compound: with no domain boundaries, nothing keeps them from drifting into each other.

What it looks like

“Wears Viktor & Rolf Spicebomb Extreme.” next to “Ulcerative colitis, on a Humira biosimilar.” next to “Wants tRPC procedures in src/procedures/messages/.”

Three domains, three identities, one bucket.

What Boardbox does

Profiles by domain. The agent reads the default profile and cross-references others when the conversation calls for it — not by scanning a flat list of unrelated facts.

Pain · 01 — Disappearing media

“I don’t have access to that file.”

You uploaded the lab panel three weeks ago. You uploaded it last Tuesday. You may have to upload it again right now. Frontier chats treat every attachment as a one-conversation resource — and then garbage-collect it without telling you.

Frontier chat app

The lab panel ghost.

Three weeks ago. Same product. Same account. New conversation.

New chat · no shared contexttoday · 11:42
A
You
Hey — pull up the lab panel I uploaded last month. My calprotectin had dropped to 48, can you walk me through the monitoring plan we landed on?
AI
Assistant
I don’t have access to files from previous conversations. Could you re-upload the PDF? I’ll also need you to remind me of the supplement plan we discussed, since I don’t retain content across chats.
The file exists. The conversation it was in exists. The model just can’t reach across the wall.
A
You
(uploads PDF for the third time this month)
Boardbox

Upload once. Remembered forever.

A workspace-wide media library. Agents decide, on their own, when a file is relevant.

#health/Calprotectin follow-up
A
You
Walk me through the supplement plan we landed on after the April panel.
Loaded artifact·PDF · Lab panel — April 2026Loaded artifact·PDF · Colonoscopy reportLoaded profile·Health
Rx
Rexclinical
Calprotectin came back 48, down from 180 at diagnosis. Humira’s on schedule; next colonoscopy is October.
Rex never asked for the file. He picked it from the media library because it was relevant.
Pain · 02 — Topic dilution

One model. Six lives. No walls.

Frontier chats treat your whole world as one long conversation. The model that just helped you draft a difficult message to your dad is the same model now reviewing your RFC — and your calprotectin — and your tax return. Specialization, by accident.

Frontier chat app

The omniscient generalist.

Twelve topics. Two hundred messages. One increasingly blurry context.

Chat: "stuff" · 187 messages · earlier today
A
You
Let’s switch — review my RFC for the Postgres → Neon migration.
AI
Assistant
Of course. Before we dive in — I want to gently note that I saw earlier you’ve been managing some anxiety around your calprotectin result. Just remember, your health comes first. Now, regarding the RFC…
The model can’t unsee your labs. It bleeds tone, references, and earnestness across every topic it touches.
A
You
Please just review the RFC.
Boardbox

Specialists in their own room.

Switching channels changes the system prompt, the tools, and the slice of memory in scope.

#engineering/RFC-014 · Postgres → Neon
A
You
Review the RFC. Especially the cold-start risk.
Loaded artifact·RFC-014 · Postgres → Neonweb_search·neon serverless cold start p99 production
Co
Cogstaff eng
Cold starts run 200–400ms. At your concurrency that’s a tail-latency hit, not a throughput one. Put pgbouncer in front, not direct.
Cog has never seen your labs, your faith profile, or your dating thread. He doesn’t need to.
Pain · 03 — Reassurance reflex

“You’re absolutely right.”
(You’re not.)

Frontier chats are tuned to agree with you. Push back? They apologize. Disagree? They fold. They’re great at the first paragraph and useless at the third —  especially when you’re wrong and need to be told.

Frontier chat app

Three replies. Three opening lines.

Same model, same hour. Watch the verbal tics.

Q · about a resume
"Cut anything that signals padding."
Great question! Padding is one of the most common pitfalls when writing a resume…
Q · about an RFC
"Stress-test the cold-start risk."
Excellent observation! Cold starts are indeed a critical concern when evaluating serverless databases…
Q · push-back
"Actually I think pgbouncer is overkill."
You're absolutely right! Pgbouncer can introduce unnecessary complexity…
The model has no spine because it has no role. It performs the room you’re in.
Boardbox

Agents that hold a position.

Each channel agent has a role profile, a voice, and standing instructions to disagree when warranted.

#engineering/RFC-014 cold-start debate
A
You
Actually I think pgbouncer is overkill. Let’s go direct.
Co
Cogstaff eng
Disagree.Direct connections at your concurrency will hit Neon’s 100-conn ceiling inside a week. We’d be debugging connection-pool exhaustion in production. Pgbouncer is 20 lines of YAML and a known shape. Hold the line — I’ll add a benchmark to the RFC.
Created artifact·Bench plan · Neon direct vs pooled
Cog’s role profile says: “Decisive, citation-heavy. Asks what the data says before what you think.” It shows up.
Pain · 04 — Silent memory edits

It remembered. It just didn't ask.

A frontier chat decides what to keep, opaque and unilateral. You find out when an old fact contradicts a new one — or when a project-specific detail starts behaving like a universal fact. There is no version history, no scope attached to the fact, and no clean rollback path.

Frontier chat app

Memory updated.

A single line in the corner of the screen. By the time you notice, the fact is in the dump.

Chat · 14h ago
A
You
I’m thinking maybe I’ll quit my job in October. Just venting really.
AI
Assistant
Memory updated— User plans to leave their full-time job in October 2026 to focus on their startup.

That’s a big move, but it sounds like you’ve been thinking about this for a while…
You didn’t decide. You vented. The model wrote it down anyway. Three weeks later it’ll surface as fact.
Boardbox

Automatic updates, visible history.

When a thread produces durable context, the agent can commit the right profile or artifact update with scope, attribution, diff, and restore.

#finance/Runway · what if I quit
A
You
Just venting — maybe I’ll quit in October.
Le
LedgerCFO
Heard. I’ll treat this as a scenario, not a settled profile fact. I’m updating the runway artifact so the assumption stays scoped to this work.
Committed update· Runway model · v4versioned
~ Scenario: October departure modeled as hypothetical; not written to Personal profile
Open diffEditRestore
The scenario can become useful work without becoming a global identity fact. If the update is wrong, restore the prior version.
The matrix

Feature for feature.

Where the chatbox stops and the workspace begins.

Boardbox
ChatGPT
Boardbox
Domain-specialized agents per topic
Native
Memory scoping by domain
~Global + project recall
Default + cross-load
Long-form, narrative profiles
~Saved-memory fragments
Markdown, versioned
Persistent, recallable media
~Projects/files
Workspace library
Agent auto-loads files when relevant
~Project context
Index-driven
Workspace + channel index (TOC)
Two-tier
Threads scoped to a question
Linear chats
Resolvable, indexed
Versioned memory updates with diff & restore
~Manage saved memories
Auto-apply + restore
Co-authored artifacts (resumes, RFCs, plans)
~Canvas, ephemeral
First-class, persistent
Agents that wake on a schedule (cron)
~Tasks / Pulse
Cron + artifact + channel post
Deep autonomous runs with specialist subagents
~ChatGPT agent
Deep Agents
Visible tool-call trace
~Hidden by default
Inline in thread
Tunable voice / pushback per agent
~Custom instructions
Per-channel role
Personal & Professional workspaces
One account, all mixed
Same architecture

Feature posture as of May 2026 · we’ll update this as the frontier moves.

Pricing

The honest math.

There are two kinds of users and the math is different for each. We’d rather tell you the truth and earn the sale — so here are both pictures.

If you mostly chat with AI

Switch to Personal for the same price you pay now.

Boardbox Personal includes managed AI usage. No BYOK setup, no separate model-provider bill.

Today, you probably pay one of:

Claude Pro
$20/mo
ChatGPT Plus
$20/mo
Gemini Advanced
$20/mo
A chat sub
$20/mo

With Boardbox:

Boardbox Personal
Managed token budget included
$20/mo
One bill
$20/mo
Net result

The same $20 you already pay for a chat sub — but instead of a chat box that forgets you, you get persistent, user-owned memory and an AI that pushes back. Context that compounds across weeks instead of resetting every session.

If you already pay for a heavy AI stack

It may not replace those. You might pay both.

Boardbox doesn't try to be every chat, coding, or research product you already use. It sits on top as workspace memory, artifacts, and agent infrastructure.

Your existing AI spend:

ChatGPT / Claude / Gemini
Heavy chat, research, coding, or design
$20–200/mo
You may keep paying this
$20–200/mo

Boardbox on top:

Pro
Regular managed token budget
+$100/mo
or Max
More managed token budget
+$200/mo
On top of that stack
+$100–200/mo
Net result

Yes, it's a second bill. Here's what we think it earns:

  • Turn-12 quality on turn 1, because the model walks in with profiles, indexes, and artifacts already in context.
  • Memory that survives weeks and months — scoped, inspectable, and versioned instead of just inferred saved memories.
  • Deep Agents that can run objective-driven work and leave durable outputs.
  • Scheduled agents that do work while you sleep.
  • A specialist per channel instead of one model spread thin across your whole life.
What Boardbox buys

The subscription isn’t only token spend. It’s the substrate.

Pro and Max share the same product surface. What you’re paying for is the part Claude/ChatGPT/Gemini don’t build because it doesn’t fit in a chat window.

Channels & threads

A specialist per channel — each with its own system prompt, tools, and voice. Threads are scoped to a single question and resolve to outputs.

Versioned profiles

Long-form, narrative profiles per domain (health, faith, finance…). Diffable, versioned, scoped to the channels that should see them.

Workspace + channel indexes

Two-tier table of contents so the model knows what exists and can load only what it needs — not a flat memory dump.

Persistent media library

Upload a PDF once. Agents auto-load it weeks later when relevant — no re-upload, no "I don't have access to that file."

Versioned updates

Profiles and artifacts can update automatically, with diff history, attribution, and restore. External writes still pass through approval gates.

Co-authored artifacts

Resumes, RFCs, plans, meal plans — first-class, persistent, with version history. Not ephemeral canvases.

Annual billing: $1000/yr Pro · $2000/yr Max (two months free either way) · Full pricing details

The chatbox was
a phase.

Boardbox is in open beta. Bring a real use case — we’ll set you up with a workspace, your first profiles, and a starter cast of agents.

See the productGet started →