MindTwin

Q: Is MindTwin only for Customer Success teams?

No. Customer Success is the first application; MindTwin itself is a governed enterprise memory platform — one ingestion pipeline, a tenant-isolated memory vault, and an intervention layer, exposed via dashboard, Slack bot, REST API, and MCP (its memory tools run inside Claude and Claude Code). Sales, support, and agent-memory use cases run on the same platform.

The renewal you lose is the one nobody saw coming.

MindTwin is a governed memory platform for enterprise AI — the layer that keeps what your organization knows current, safe, and actionable. Its first application is Customer Success: it ingests every account conversation across Slack, email, and calls — then does three things no search tool does: flags renewal risk on its own when someone's stance quietly reverses, vets outgoing messages against the record before a tone-deaf email costs you trust, and governs sensitive content at write time so PII and crisis disclosures never become retrievable. Under the hood: biologically grounded memory that fades, strengthens, and consolidates — so the current truth outranks the stale one.

Patent Pending

Run a 6-Week Pilot Explore the Science

Published research · arXiv:2609.28575 Open benchmark, data & results 180+ automated tests Per-team isolation (Postgres RLS)

It caught a real one

From a live workspace: the account's champion called the renewal “a no-brainer” in February. By June he was evaluating a competitor. MindTwin's inner voice connected the two statements months apart and raised the risk — before anyone asked.

MindTwin Unresolved tensions dashboard flagging a stance reversal on a renewal, with severity and the dated prior belief it contradicts

Actual dashboard — the tension MindTwin raised on its own, with the dated prior belief it contradicts. One click resolves it; the refuted belief then fades faster.

Download the product overview (PDF)

We also built and open-sourced TWIST — a benchmark for whether AI memory intervenes correctly when beliefs change, now a published paper. Paper (arXiv) → · GitHub → · The story →

Why Enterprise AI Memory — and Why Now

2026 is the year every AI assistant got a memory. The enterprise question has already moved on: not “can it remember?” but “can you trust what it remembers — and what it does about it?”

1 · The memory layer is inevitable

Every serious AI deployment is getting persistent memory — the major platforms shipped it, an ecosystem of vendors sells it, and a wave of 2026 research now benchmarks it. Meanwhile your organization's actual memory already exists — scattered across Slack, email, and calls — and today it mostly evaporates: every departure, handoff, and re-org is organizational amnesia, re-purchased at payroll prices. The knowledge either compounds in a memory layer or walks out the door. We've written about the scale of this →

2 · Remembering is a commodity. Being right is not.

Storage plus similarity search — “just add RAG” — is now table stakes, and it carries failure modes an enterprise cannot live with. Beliefs change: a memory that asserts January's truth in June is worse than no memory, because it's confidently wrong. Assistants that over-flag get muted within a week. Sensitive disclosures, once stored, become permanently retrievable. And all three problems get worse as the memory grows — we measured it. The hard problems in enterprise memory are currency, restraint, and governance. Nobody buys a system of record they can't trust.

3 · We build the layer that acts on truth — and we published the yardstick

MindTwin is memory that intervenes: it flags stance reversals unprompted with dated evidence, vets outgoing messages before they cost you trust, retires refuted beliefs while preserving the history, and governs PII and crisis content at write time. We didn't just claim this category — we defined how to measure it: TWIST, the published benchmark for intervention quality, where we report every system's scores including our own failures. Customer Success renewals are the beachhead because the money is most visible there — the same layer serves sales, support, and, via MCP, the agents your company is deploying right now.

The Business Math

Conservative worked model — swap in your own numbers in a pilot. MindTwin pays for itself if it saves one at-risk renewal a year; everything else is upside.

~$36k

retained ARR per CSM per year — 30 accounts × $120k ARR, 12% churn, a third late-detected, a quarter of that caught in time. Deliberately conservative.

1,840 hrs

returned to a 10-CSM team yearly — 4 hrs/week each of context-hunting across Slack, Gong, and email replaced by catch-up briefings (≈$116k in recovered capacity)

1 renewal

a single saved Acme-sized renewal ($180k) dwarfs the annual subscription on its own

The two assumptions that matter — how much of your churn is late-detected, and how much of it gets caught in time — are exactly what a 6-week pilot measures on your own accounts. That's the point of the pilot.

What MindTwin Actually Does

Four jobs, all live in the product today — each one built on the cognitive engine underneath, each one measured in public.

🚨

Catches the Quiet Reversal

Your champion said “renewing is a no-brainer” in February. In June he's “evaluating options.” MindTwin's inner voice connects statements months apart — across Slack, email, and calls — and raises the tension unprompted, with the dated evidence attached.

🛡️

Stops the Bad Email

Before a draft goes out, MindTwin checks it against everything the account actually said. And it doesn't cry wolf: on our own published benchmark it almost never flags a safe message — restraint is a feature you can measure.

⚡

Handoffs Without the Archaeology

A rep leaves, an account moves — the next person gets a catch-up briefing built from the account's living memory: current beliefs first, resolved history preserved underneath. CSMs typically burn ~4 hours a week hunting for this context.

🔒

Governance at Write Time

PII is scrubbed and crisis disclosures are suppressed before storage — not filtered at read time — with a content-free audit trail. Zero false suppressions across 5,882 emotionally varied turns, and legitimate emotional context stays recallable.

The engine underneath is biologically grounded — ACT-R-style activation, Ebbinghaus-style decay, sleep-cycle consolidation of episodes into durable account traits — because “remember everything forever” is exactly how memory systems end up asserting stale beliefs. Forgetting the refuted is the feature.

How the Platform Works

One write path. Every message from every source passes through governance, weighing, and belief-tracking before it becomes memory — that's what makes the memory trustworthy.

Ingest

Slack, email, and call transcripts flow through a single pipeline — same rules for every source

→

Govern

Write-time safety gate: PII scrubbed, crisis content suppressed, content-free audit trail — before storage, not filtered at read time

→

Weigh

Salience scoring routes each memory to a tier — episodic, semantic, or core — each with its own decay profile

→

Distill

Facts extracted with provenance — every belief carries receipts back to the exact turns it came from

→

Watch

The inner voice compares new statements against held beliefs; contradictions land in a tension ledger with dated evidence

→

Act

Unprompted risk flags, catch-up briefings, draft vetting — and once a tension is resolved, the refuted belief decays faster

One Memory Layer, Every Team That Lives on Relationships

Customer Success is the beachhead — the money is most visible where renewals slip. The platform underneath is general.

Customer Success · live today

Renewal-risk tensions, draft vetting, handoff briefings — the full workflow this page describes, running as a product.

Sales & Account Teams

The same engine tracks deal truth: what the buyer actually committed to, where their stance moved, what your follow-up is about to get wrong. Design-partner territory today.

Support & Ops

Context that survives ticket handoffs and shift changes — current state first, history preserved, sensitive content governed. Same vault, different surface.

Your AI Agents · via MCP, today

Every agent your company deploys needs memory it can trust. MindTwin's memory tools already run inside Claude and Claude Code via MCP — governed recall, belief currency, and write-time safety as an API, not a promise.

Platform Architecture

An application for teams today; a governed memory API for your agents tomorrow — same vault, same write-time governance, same tenant isolation.

Sources

SlackGmailGong (call transcripts)

↓

Ingestion + Cognitive Middleware — one write path

Write-time safety gate: PII scrub · crisis suppression · content-free auditSalience scoringFact extraction with provenance

↓

Memory Vault

Episodic (high decay)Semantic facts (low decay)Core (near-zero)Activation re-rankingSleep-cycle consolidationSupersession: refuted beliefs decay faster

Per-team isolation — Postgres row-level security; a cross-tenant test suite gates every change.

↓

Intervention Layer

Tension ledgerDraft vetting (check_alignment)Catch-up briefingsUnprompted risk flags

↓

Surfaces

DashboardSlack botREST APIMCP — memory tools inside Claude & your agents

The Science

Built on 50+ years of ACT-R cognitive architecture research.

Cognitive Activation Model

Based on ACT-R Theory

Our proprietary activation model, grounded in 50+ years of ACT-R cognitive architecture research, determines which memories surface. Frequently accessed, recent memories stay vivid. Unused ones fade — just like the human brain.

Salience Detection

Multi-Signal Scoring

Emotional intensity, explicit intent, and semantic novelty combine in a proprietary scoring model to automatically flag potential core memories for human validation.

Proven, Not Promised

Every claim below is checkable — the benchmark harness, methodology, and test suite ship with the product. The memory tools also run inside Claude and Claude Code today via MCP.

1,540

questions — evaluated on the full LoCoMo long-conversation memory benchmark, reproducible with one command

crisis false-positives from the safety gate across 5,882 emotionally varied conversation turns

180+

automated tests, including a mandatory cross-tenant isolation suite that gates every change

system configurations evaluated on TWIST, our published benchmark (arXiv:2609.28575) — none pass yet, ours included, and every per-item output is public. That's what “proven, not promised” means.

Frequently Asked Questions

What is MindTwin?

MindTwin is a governed memory platform for enterprise AI — the layer that keeps what an organization knows current, safe, and actionable. Its first application is Customer Success: it connects to Slack, Gmail, and Gong, remembers every account, flags renewal risk unprompted when a customer's stance changes (with citations), vets outgoing drafts against the record, and governs PII and crisis content at write time. The same platform serves sales, support, and AI agents via MCP.

How is MindTwin different from RAG or enterprise search tools?

Search and RAG tools retrieve what was said. MindTwin also notices when what is being said changes: its inner voice detects contradictions and stance reversals across months, vets outgoing drafts against the record before they're sent, and lets refuted beliefs fade via processing-aware decay.

Which tools does MindTwin connect to?

Slack, email (Gmail), and call transcripts (Gong). Pilots are white-glove: we wire the connectors up with your team rather than handing you a self-serve flow — deliberate at this stage, so ingestion quality is verified before the memory goes live. Memory is per-team, enforced with Postgres row-level security, so a team can only ever see its own memory.

Is customer data safe in MindTwin?

PII such as emails and phone numbers is redacted before anything is stored, crisis content is never stored (the bot responds with a helpline referral instead), and every suppression is auditable without storing content. In benchmark runs the safety gate produced zero false positives across 5,882 conversation turns.

Is MindTwin only for Customer Success teams?

No — Customer Success is the first application because renewal risk makes the value most measurable. MindTwin itself is a governed enterprise memory platform: one ingestion pipeline, one tenant-isolated vault, one intervention layer, exposed through a dashboard, a Slack bot, a REST API, and MCP (its memory tools already run inside Claude and Claude Code). Sales, support, and agent-memory use cases run on the same platform and are open for design partners.

What is the TWIST benchmark?

TWIST is an open benchmark created by MindTwin for intervention quality in conversational memory: unprompted tension detection, output-time draft vetting, belief supersession, and safe recall — each paired with controls that price false intervention. It extends LoCoMo and is published as a paper (arXiv:2609.28575); the human-validated dataset, harness, and all per-item results are public at github.com/subratpanda/twist-benchmark.

How does a MindTwin pilot work?

Six weeks on one at-risk segment, fixed fee scoped on one call, one success metric agreed up front — for example, surface at least three at-risk renewals the team would have missed. You keep the findings either way.