Data terms

Training and retention comparison
for teams

What no training does and does not cover, which products document retention periods and sub-processors, and the evidence to ask for before sensitive work moves

Aug 18, 2026 · 13 min read

The short version
No training is the first control

A team that needs AI which does not train on its content should buy a business workspace with an explicit no-training commitment, documented retention controls, a current sub-processor list and evidence an administrator can see in the product. The commitment has to cover the workspace vendor and every model provider that receives a prompt. The eight workspaces compared on the same criteria below are Playgram, WorkLLM, nexos.ai, Langdock, TeamAI, Aymo, Magai and TypingMind.

The phrase itself is easy to publish and hard to verify, and it answers only one of five questions. The others are who receives the data, where it is processed, how long each copy remains, what memory derives from the original content, and what an administrator can restrict, inspect, export and delete.

This guide sets out what the eight products publish on each of those, and where a marketing claim outruns the documentation. It ends with the evidence to collect before sensitive work moves in, and a controlled test that checks deletion rather than trusting it.

Who this guide is for
Which teams this fits

IT & security01

IT and security owners

You are replacing unapproved consumer accounts and need evidence rather than assurances.

Legal02

Legal and compliance

You read the terms and the privacy page and have to reconcile the two before signing.

Restricted03

Teams with restricted material

Roadmaps, interview notes, client files and HR documents pass through prompts every week.

Not yet04

Teams that do not need this

One or two people working only with public information on a single approved product.

The real problem
Why one policy cannot cover four tools

Each subscription creates its own privacy boundary, retention rule and administrative surface, and nobody is looking at all of them together.

01

Cost

Separate business plans duplicate model access and charge for provisioned people whether or not they use them. As an example, four single-vendor team plans came to about $101 per person a month at July 2026 list prices, taking ChatGPT Business at $25, Claude Team at $25, Gemini Business at $21 and Grok Business at $301234. Read it as an illustration rather than a rate, because a cheaper mix is easy to assemble. On this topic the review effort matters as much as the fee, since each product adds a retention policy, a sub-processor register and a set of terms for somebody to read. Idle seats are also hardest to find when departments buy their own, and each provider defines usage and minimum members differently.

02

Workflow

People switch tabs, copy prompts and download then upload the same file again, so one sensitive document passes through several vendors in an afternoon. Two risks follow. Nobody can reliably remember which account carries the approved business terms, and moving a task between models leaves several stored copies of the same material behind.

03

Context

Chat history is usually private to one account, so instructions, decisions and corrected facts never become team knowledge. The team then repeats the same client brief, policy or project background in each tool, which costs time and creates more copies of material that may be sensitive. Memory makes that worse rather than better unless its scope is clear, because a stored summary or extracted fact is another copy under another rule.

04

Management

Consumer accounts leave a manager unable to verify which model received a prompt, whether the work happened in a business workspace, or whether history and training controls were on. Nobody can say which integrations reached the conversation either, or whether a person's content was deleted after they left. A business workspace helps only when it shows those facts to an administrator, because a policy saying your data is protected is not operational evidence.

What to look for
The controls behind the promise

Five checks that turn a policy statement into something an administrator can show an auditor. The fourth is where most products in this group stop.

Coverage

Approved models under one policy

Approved OpenAI, Anthropic, Google, xAI and open-model endpoints should sit behind one controlled interface, and the workspace has to use business or API terms that stop providers training on prompts and outputs. An admin should be able to disable a model whose region or retention fails policy.

Tools

The tools that touch the data

Check each product for web search and cited research, document drafting, spreadsheet analysis, image generation, video generation, code review or execution, chats that leave no history, and MCP or API connections. Every one of those is another path your content can take out of the workspace.

Memory

Retention split by layer

Workspace history, backups, memory databases, application logs and model-provider retention are five different periods, and most vendors publish one. Ask for all five, and look for a retention control an admin can set rather than a paragraph describing a general practice.

Control

Evidence an admin can produce

Logs should identify person, model, period and ideally the project, and deletion should appear as an event rather than as a promise. Add a current sub-processor register with change notices, permissions that follow user and group, and workspace ownership that survives somebody leaving.

Pricing

Pre-bill limits and a quiet mode

Hard caps or automatic blocking should act before unapproved spending happens rather than appear in a report afterwards. Teams handling sensitive one-off work should also look for a documented mode that leaves no history or long-term memory, which is the clean answer for a task nobody should retain.

The shortlist
What each product covers and costs

The multi-model workspaces a team is most likely to weigh up, on the same criteria and to one standard. Where a vendor does not document something, the cell says so.

Product
Best for
Model access
Pricing
Shared team memory
Cross-model context
Notes
Playgram
Teams wanting no training with a usage-based team plan
The latest GPT, Claude, Gemini and Grok models and many more30
Credits, with no per-seat fee: $60/mo for 10,000 credits billed monthly, so five people pay the same $6029
Yes, at team, project and personal scopes30
Yes, switch mid-thread and the conversation carries over30
Video generation is not shipped yet, and MCP is not publicly documented30
WorkLLM
Teams that want zero retention stated at the model layer
More than 200 models6
Per seat: Basic $20 per user/mo billed monthly with 2,000 pooled credits per user, so five users pay $1006
Organisation memory applied automatically, and agents can also run without it7
Manual test required
Internal workspace retention, named model providers, countries and a public sub-processor list were not found8
nexos.ai
Teams that need a named sub-processor list with locations
More than 200 models10
$39/mo for the 1-month AI Workspace plan with 1,000 credits. The page does not state included seats, so a five-person total is unverified10
Projects keep uploads, searches, conversations and instructions for their members11
Yes, project context stays in place when the model changes11
The privacy matrix marks some coverage as partial, and item-level memory correction and expiry are not documented32
Langdock
Teams that want retention as an admin setting
Claude, GPT, Gemini and others, included on the Business plan13
Per seat: Business EUR 25 per user/mo excluding VAT, so five users pay EUR 125. Business Max is EUR 99 per user/mo13
No. Automatic memory is personal, capped at 50 entries and unavailable in project or agent chats37
Yes, a model can be changed mid-conversation while the thread is kept37
No automatic team memory, video generation or documented no-trace chat14
TeamAI
Teams comfortable reconciling the terms with the privacy page
Hosted models from OpenAI, Anthropic, Google, Meta and DeepSeek among others17
Per workspace: Professional $149/mo for up to 25 users with 20,000 credits, so five users also pay $14917
No. Automatic memory is personal and persists until it is disabled or reset38
Manual test required
The terms grant a broad perpetual licence over submitted and generated data, and no full sub-processor list was found34
Aymo
Small teams whose work stays inside a provider's own terms
More than 50 models across GPT, Claude, Gemini, Grok, DeepSeek, Qwen and Mistral21
Per workspace: Premium $20/mo for up to 10 members, so five users pay $20. Business is $39/mo and holds 2521
Team memory is advertised, and whether entries are saved automatically or can be inspected is not documented35
Yes, changing model does not start a new chat22
Conversation history remains until deletion or account closure, and no full sub-processor list was found35
Magai
Creative teams wanting a stated backup purge window
Every model and tool inside one usage balance, drawn down at different rates23
Per seat: Standard $20/mo plus $20 for each added member, so five users pay $10023
No. Chats are private by default and become shared only when somebody shares them36
Manual test required
Server and application logs carry no stated end date, and MCP, code review and no-trace chat are not documented36
TypingMind
Teams that want the workspace store on their own servers
GPT, Claude, Gemini and custom models through keys an admin provides25
Per workspace: Starter $99/mo with five seats included, then $8 per extra seat. Provider API charges are separate25
No. Shared prompts, agents and knowledge bases are configured by hand25
Manual test required
Retention is described as for as long as necessary rather than as a fixed period37

This table compares multi-model team workspaces with each other. The single-vendor plans a workspace replaces are priced further down, under 'Priced per seat', and are not rows here. Pricing is the lowest-priced paid plan that covers five users, at the monthly rate, so a product whose entry plan holds fewer than five people is shown on the plan that holds them. Each cell cites the page that documents that cell rather than one pricing page per row. Figures checked August 2026, and cells marked 'Manual test required' could not be confirmed from public documentation.

Controls and data
Retention regions and evidence

The same products on the criteria this topic turns on: what each one does besides chat, what it connects to, what an admin can see and cap, and where your content is processed and kept.

Product
Built-in tools
Integrations
Usage visibility
Usage controls
Training on your data
Where the models run
Playgram
Image generation, web search, deep research, document and spreadsheet work, and code execution30
Not publicly documented30
Adoption, query volume and model preference by person30
A credit limit per person, a limit across the whole team, and model access set per user29
No31
US-based infrastructure, with a choice of US or EU data residency on the plan pages29
WorkLLM
Shared threads, web-oriented research, files and agents6
Named work-app connectors, while MCP and a public API are not documented9
Reports with role-based access and audit logs, and the exact fields are not public6
Model and data usage controls are claimed, and hard-budget behaviour is not documented6
No, with zero retention stated at the model layer and an isolated tenant per customer8
Not publicly documented, and no sub-processor list was found8
nexos.ai
Web research, document and slide generation, charts and agents12
MCP integrations and an API gateway12
Metrics by model, user and request type, with logs12
Hard budget caps documented at gateway and enterprise level12
No by default, and zero retention at the model layer can be enabled32
A dated sub-processor list names AWS and Cloudflare in the EEA, plus optional model providers in the EEA, the US and other regions32
Langdock
Image generation, deep research, web search, document editing and spreadsheet analysis14
MCP, APIs, Excel and Outlook add-ins and further connectors15
Active users, messages and models over a reporting period an admin chooses37
Admins can disable models and set workspace and user spend limits37
No, and models whose provider retains data longer are labelled, with one example retained up to 60 days for abuse prevention33
Multi-tenant on Azure in the EU, with most model hosting configurable for the EU33
TeamAI
Web search, research mode, code and data analysis, documents and workflows20
An MCP server and further connectors17
Usage reporting, with person by model by period detail not publicly described19
Spend limits block overage by default, and an admin can opt into a hard dollar cap with notifications19
TeamAI states that neither it nor its providers train on prompts, uploads or outputs34
Personal information is processed in the US and retained as necessary, with no fixed content-retention period34
Aymo
File context, uploads, web search and deep research22
Your own provider keys, while a public API and MCP are not documented22
Not publicly documented at a person or model level22
Pre-bill administrator caps are not documented22
Private conversations are not used to train Aymo's own models, and prompts are handled under the selected provider's terms35
Not documented beyond a named payment processor, with history kept until deletion or account closure35
Magai
Image and video generation, shared libraries and real-time collaboration23
More than 100 integrations, while MCP and a developer API are not confirmed24
An admin dashboard covering usage, seats and billing23
Pre-bill per-user or per-model hard caps are not documented23
No, and Magai says it selects API providers that contractually do not train on API data36
Servers are in the US, content stays for the life of the account and is purged from backups within seven days of deletion36
TypingMind
Web search, browsing, image generation, plugins and knowledge bases25
Plugins, API integrations and knowledge connectors27
Analytics and chat logs for admins on the higher plans25
Model usage limits can be imposed by user or group25
Cloud-hosted team data sits on TypingMind servers, and self-hosted data sits in your own database37
The processing agreement names Render, Vercel, AWS, OpenAI, Sentry, Customer.io, Stripe and Lemon Squeezy with processing countries37

These criteria decide daily use more than the model list does, and vendors document them very unevenly. 'Not publicly documented' means the official sources checked did not state it, and it does not mean the feature is absent, so read those cells as questions to put to the vendor. Public documentation is also not a substitute for the signed processing agreement and provider schedule. Checked August 2026.

Priced per seat
What the single-vendor plans cost

The published per-seat price of each major single-vendor team plan, billed monthly. Each one also brings its own retention policy and its own sub-processor list to review.

Provider
Plan
Per seat
Models
ChatGPT Business
Business · billed monthly ($20 billed annually)
$25/seat/mo
GPT family (GPT-5 Instant, Thinking) + o-series reasoning models
Claude Team
Team (Standard seat) · billed monthly ($20 billed annually); 5-seat minimum
$25/seat/mo
Full Claude model family (Sonnet, Opus, Haiku)
Gemini Enterprise (Business)
Gemini Enterprise, Business edition · annual commitment (Standard is $30 with commitment)
$21/seat/mo
Gemini via the Gemini Enterprise app
Grok Business
Grok Business · billed monthly, no published annual discount
$30/seat/mo
Grok family (Grok 4, Grok Heavy)

Buying all four for one person came to about $101 a month at July 2026 list prices. Read that as one example stack rather than a going rate, because a team can assemble a cheaper mix and the four plans do not buy the same amount of use. On this topic the review effort matters as much as the price, since four products mean four retention policies, four sub-processor registers and four sets of terms for a legal team to read. Figures checked July 2026.

The cost drivers
What the review itself costs you

Cost is the secondary question on this topic, and two of these are paid in legal and security hours rather than in subscription fees.

Review per vendor

Every product added means another retention policy, another sub-processor register, another processing agreement and another security review. That work repeats whenever a vendor updates its terms, and the same legal and security people do it each time.

Plans per person

Each plan is priced per person, so a second one multiplies the total rather than adding a line. Buying all four of the plans priced above came to about $101 per person a month at July 2026 list prices1234.

Stored copies

Moving one task between products creates several stored copies of the same document, each under a different retention rule. Nobody is billed for that directly, and it becomes expensive the moment somebody has to prove where a file went.

Idle seats

Seats bought by a department or an individual are hard to spot, and each provider counts usage and minimum members differently. Claude Team bills a five-member minimum whether or not five people are working2.

The options
Six ways to control where data goes

Six setups, ordered by how much an administrator can verify rather than by how strongly each one is described.

A governed multi-model workspace

One vendor routes requests to several providers under one policy, one set of logs and one contract. That vendor becomes an additional processor, so it has to prove what it sends onward, what it stores and what it logs.

Best for: Teams with non-public work across more than one model.

Strengths

  • One retention policy and one sub-processor register to review rather than four
  • An admin can restrict models whose region or retention fails policy
  • Usage and deletion evidence sit in one place for an audit

Trade-offs

  • The workspace adds a processor between you and the model providers
  • Several vendors publish strong claims without naming sub-processors or retention periods
  • A no-training promise on a marketing page can still sit beside a broad content licence in the terms

A workspace with retention controls

The same setup where an admin chooses how long content is kept and which models may be used. Langdock documents inactive-chat retention of seven days, one month, three months, twelve months or permanent, and flags models whose provider keeps data longer33.

Best for: Teams with a data-classification policy to enforce.

Strengths

  • Retention becomes a setting with an owner rather than a paragraph in a policy
  • Models with longer provider retention can be labelled or disabled before anyone uses them
  • Deletion can be tested against a stated period rather than a general promise

Trade-offs

  • Workspace retention and model-provider retention are separate, so both have to be checked
  • Shorter retention removes the history and project context a team may rely on
  • A control on the admin page still needs a test account to confirm it reaches search and memory

One provider for the whole team

Standardise on a single vendor's business plan and review one set of terms. Governance is simpler because there is one policy, one console and one processor to name in a register.

Best for: Teams whose work fits one approved ecosystem.

Strengths

  • One security review, one processing agreement and one region policy
  • The vendor's own controls are usually deeper than a third party can wrap

Trade-offs

  • No independent model comparison, so a workflow that vendor handles badly moves off-policy
  • Work outside that ecosystem tends to reappear in personal accounts nobody can see

Several business plans

Buy each vendor's business product for the native controls. The capability is strong and the review effort multiplies, because four products mean four retention policies, four registers and four consoles.

Best for: Regulated teams that need each vendor's own terms.

Strengths

  • Each vendor's own security features and support are available
  • Contract terms can be negotiated with each provider directly

Trade-offs

  • Four retention and sub-processor reviews, repeated whenever a vendor updates them
  • The same sensitive document can end up stored in several products at once

Self-hosting the workspace

Run the product on your own infrastructure so the workspace data stays in a database you control. TypingMind documents self-hosted team deployments where workspace data sits in the customer's own MySQL database, while cloud-hosted plans store it on the vendor's servers37.

Best for: Teams with infrastructure requirements written into contracts.

Strengths

  • The workspace store is yours, which removes one processor from the chain
  • Region and backup policy follow your own infrastructure rules

Trade-offs

  • Model calls still leave for whichever providers you configure, so provider terms remain
  • Somebody has to run upgrades, backups and incident response for it

A custom API build

Engineers select the endpoints, databases, logs and deletion jobs directly. It gives the most control over where content goes, and the team then owns authentication, the interface, routing, monitoring, retrieval security and incident response.

Best for: Regulated or technical teams with engineering capacity.

Strengths

  • Every storage location and deletion job is a decision you made
  • Provider terms can be chosen per endpoint rather than accepted as a bundle

Trade-offs

  • Retrieval security and log hygiene become your problem, and both are easy to get wrong
  • The build competes for the same engineering time as the product itself

In practice
How restricted work moves safely

An unreleased launch plan in a restricted project, where the members are named, the classification is recorded and the last step is a deletion check rather than a handover.

Restricted project - launch brief, research notes, brand rules, data classification Analysis a long-context model risks and dependencies Human review personal data removed assumptions marked Challenge a second model approved context only Saved at project level drafts deleted after export retention checked

A person removes personal data and marks unsupported assumptions before the work moves on, and sends anything thin back to the analysis stage. The approved output is stored at project level rather than in an unrestricted company store, and closing the project means exporting the result, deleting the drafts and confirming that the retention job covers chats, files, indexes and backups.

Shared memory
Treat it as retained data

Memory is a storage system with a friendlier name, so on this topic it belongs in the retention review rather than in the convenience column.

Definition01

Memory is not the context window

A context window is the information placed into one request, and it creates no durable knowledge by itself. Chat history is a stored transcript that nothing retrieves for you. Memory is stored outside the window and pulled into later work, which makes it a data store with its own retention rule.

Shapes02

Products build it four ways

WorkLLM documents organisation-level memory and nexos.ai keeps shared project context. Langdock and TeamAI both keep automatic memory personal to one account. Magai and TypingMind rely on knowledge bases somebody builds, and Aymo advertises team memory without its controls353738.

Scope03

Ask what is actually stored

Find out whether an entry is a verbatim message, a generated summary, an embedding or an extracted fact, and which model providers receive it when it is retrieved. Then ask whether project memory is isolated from other projects and whether a sensitive session can bypass memory and history entirely.

Control04

Deletion is the test that matters

Check that users can inspect and correct an entry, that an admin can delete one, and that deleting a chat also removes anything derived from it. TeamAI notes that turning memory off can take up to 24 hours to stop influencing responses, which is the kind of delay to find before it matters38.

A two-week trial
How to verify the data terms

A vendor-neutral plan that collects evidence rather than assurances, and ends by checking that deleted content is really gone.

01

Collect the documents

Ask each candidate for the processing agreement naming it as processor, the current sub-processor register with a way to be notified of changes, and a model inventory showing provider, region and retention class. A vendor that cannot produce those three for the plan you would buy has answered the question already.

02

Separate the retention layers

Write down four periods for each product: workspace history, backups, application logs and model-provider retention. They are different numbers and vendors often publish only one. Note which models carry longer provider retention, since some are flagged for abuse review rather than for training.

03

Run restricted work through it

Set up a project with real but classified material, invite only the people who should see it, and run the work across two models with a review stage between them. Check that the second model received the approved project context rather than a fresh upload, and that nothing reached an unrestricted store.

04

Test deletion for real

Delete a chat, a file and a stored memory entry, then search for them in history, in project search and in the memory interface. Ask whether deleting a source chat also removes anything derived from it, and record how long each disappearance took rather than accepting a policy statement about it.

05

Check what an admin can prove

Confirm usage exports by person, model and period, deletion events in an audit log, model restrictions, and budgets that block before an overage. Then reconcile the terms against the privacy page, because a no-training statement and a broad content licence can sit in the same contract.

Bottom line
Ask for evidence not assurances

A team can answer the no-training question by moving approved work into a governed business workspace, and that phrase is only the first control. A credible setup also shows who receives the data, where it is processed, how long each copy remains, what memory derives from the original content, and what an administrator can restrict, inspect, export and delete.

The public evidence is uneven, and it is worth saying which way. nexos.ai publishes a sub-processor register with locations. Langdock documents practical retention periods and labels models whose provider keeps data longer. TypingMind offers the strongest architectural control through self-hosting while its cloud plan adds vendor-side storage. WorkLLM states zero retention at the model layer and publishes less about its own. TeamAI's broad content licence needs legal review, Aymo's provider commitment is narrower than its headline, and Magai gives a purge window while its logs have no stated end.

None of that is a ranking, and none of it replaces a signed agreement you have read yourself. Plans cannot be compared by name, public pages are not the processing agreement, and retention policies change. So run the work through a controlled test with your own material, then take every unsupported or contradictory point to the contract.

The right buy
When it fits and when it does not

Not the right buy when

  • Work that only ever involves public information
  • A contract that requires the system on infrastructure you control
  • One approved single-vendor platform that already covers the work

The right buy when

  • Non-public material passes through prompts most weeks
  • Several models are in use and only one policy should apply to them
  • An admin has to show who used which model and what was deleted

Where Playgram fits
And where it does not

Two questions settle this one: how much of the work involves material that should not leave a controlled boundary, and whether an administrator could prove today where any of it went.

If restricted material is routine and nobody can prove that, you are shopping for evidence rather than for a stronger promise. A product there has to state the no-training position for the plan you would buy, publish or provide a current sub-processor list, separate workspace retention from model-provider retention, and show an admin who used which model and when. Ask for all four in writing and test deletion yourself.

If the team works only with public information on one approved product, this whole review is more than the job needs. A contract that dictates where data is stored usually points to self-hosting or a build of your own, because those are the only shapes that remove a processor rather than adding one.

Playgram belongs on the shortlist beside the others here for the first case, which is restricted work across more than one model with context that has to stay inside the boundary. The memory part of it is covered by the three scopes below, so read those first, then run the estimator with your own numbers.

Team memory

Shared across everyone and every model.

Project memory

Scoped to a campaign or document set.

Personal memory

Your own working style, kept private.

Fair pricing
Pay per usage, not per seat

Upgrade as needed, and only pay for what you actually use

Save ~17% with the annual plan

Pro

$50/ month

Perfect for small and medium teams

Unlimited users & infinite memory

Multi-LLM chats

Granular access control to models

EU data residency

Get started

Ultra

$200/ month

Best for large, growing teams

Unlimited users & infinite memory

Multi-LLM chats

Unlimited use of DeepSeek V4 Flash

Granular access control to models

Choose US or EU data residency

Get started

Enterprise

Get in touch

Unlimited Credits

For organizations with advanced needs

Unlimited users & SSO

Priority Support

Unlimited use of DeepSeek V4 Flash

Granular access control to models

Choose US or EU data residency

Book a call

30-days money back guarantee

Pricing Calculator

Team size
people
Usage per person
messages/day
Usage complexity
Docs, coding help
Auto mode
%

Playgram will automatically choose the most cost-efficient model suitable for the task. It will be chosen by users in approximately 80% of requests. Your models for the remaining 20%:

If you bought each separately:

ChatGPT Business$800 / month
Claude Team$1 760 / month
Gemini Business$840 / month
Grok Business$1 200 / month
Total$4 600 / month

Playgram

$300/ month

~59 000 credits / month · ~$8 / user

Save ~$4 300 / month
Get started

Frequently asked
questions

No, and conflating the two is the most common mistake in this review. Training is what a provider may do with your content to improve a model. Storage is how long the workspace and its providers keep a copy so history, projects and memory can work. WorkLLM states zero retention at the model layer, while Langdock, TeamAI, Aymo, Magai and Playgram all store some workspace content by design, so ask about both separately[6][33][35][36].

It depends on the product, and only some of them publish a period. Langdock lets an admin choose inactive-chat retention of seven days, one month, three months, twelve months or permanent storage[33]. Magai keeps content for the life of the account and purges it from database backups within seven days of deletion, while its server logs have no stated end date[36]. TeamAI and TypingMind both describe retention as being for as long as necessary rather than as a fixed period[34][37].

Fewer than you would expect, and it is the fastest way to separate a policy from a practice. nexos.ai publishes a dated list naming AWS and Cloudflare in the EEA alongside optional model providers across the EEA, the US and other regions[32]. TypingMind's data-processing agreement names Render, Vercel, AWS, OpenAI, Sentry, Customer.io, Stripe and Lemon Squeezy with processing countries[37]. Langdock points to a trust centre, and several others name only a payment processor.

The terms rather than the privacy page, because the two can disagree. TeamAI states that neither it nor its model providers train on prompts, uploads or outputs, while its terms also grant a broad perpetual licence over submitted and generated data, and those two need reconciling before signature[34]. Aymo says private conversations are not used to train its own models and that prompts are handled under the selected provider's terms, which is narrower than an unconditional commitment[35].

Ask, because the answer is rarely on a marketing page and it is the question memory makes necessary. Find out whether the stored item is a verbatim message, a summary, an embedding or an extracted fact, who can inspect and correct it, and whether an administrator can delete it. TeamAI notes that disabling memory can take up to 24 hours to stop influencing responses, which is the kind of detail worth having in writing[38].

Eight things, and each of them is checkable rather than promised. Ask for a current sub-processor register with change notifications, a model inventory showing provider and region, a retention control with selectable periods, and deletion events in an audit log. Then ask for usage exports by person, model and period, hard budgets or automatic blocking, a data-processing agreement naming the vendor as processor, and a test account proving that deleted content leaves search, history and memory.

Related comparisons

Shared AI memory for teamsAI tools for agencies and client workConsolidate AI subscriptions

Stop paying per seat
Give the whole team every model

Every model, shared team memory, one team plan priced by usage not per seat

Create workspaceEstimate your bill