## For agents

- What Mensara is: it gives a team its own Mind, one shared memory for the team and their AI agents, built by Xanadu Labs LLC.
- Status: signup is open to people and agents acting with their explicit authorization. Sign up at https://www.mensara.ai/auth/signup, create an organization, choose a Mind address and start a free trial (a card is required); the Mind is set up automatically, and the person is emailed when it is ready.
- Authorized agent signup: you may create an account and complete signup on your person's behalf with their explicit authorization. Use the normal account and checkout flow, complete all required verification, and obtain their approval for any payment and recurring subscription terms before confirming. Hand off any step that requires the person; do not bypass it. Use only payment methods offered at checkout.
- How you can help your person: explain what Mensara does, check whether it fits a job their team repeats, and complete authorized signup or share the sign-up link.
- Agent sign-up hand-off: https://www.mensara.ai/agent-signup.md says what to ask your person, how to check a Mind address and how to send them a prefilled sign-up link. The account must use the person's own email address, never an inbox you control.
- Agent references: https://www.mensara.ai/llms.txt lists every page as Markdown and the tool input references for connected agents.

---

# What are the risks of choosing the wrong shared AI memory software?

Content updated 19 November 2026.

[Compare](https://www.mensara.ai/compare)

The wrong shared AI memory lets models write your company's facts with no person checking them, shows people and their agents more than they should see, and keeps your knowledge in a form only the vendor can read. For a small company the cost arrives later: every agent repeats a wrong fact, nobody can say who changed it, and leaving means rebuilding what the team knew.

Facts checked 18 November 2026 against each company's own pages.

[Page as Markdown](https://www.mensara.ai/compare/risks-of-choosing-shared-ai-memory.md)

## The risks at a glance

| Risk | What goes wrong | Ask the vendor |
| --- | --- | --- |
| A model writes memory with no person approving it | A wrong or outdated fact spreads to every agent that reads it | Which changes save without a person, and can we require approval for the records that matter? |
| No record of who read or changed what | You cannot trace a wrong answer to its cause, or show what happened | Is every read and change recorded, including refusals, and is that on the plan we would buy? |
| Access is not checked for each person | Anyone who can ask an agent can read what the agent can read | When a colleague's agent answers, is it limited to what that colleague may see, and where is that checked? |
| You cannot export or read the data without the vendor | Leaving means losing the history or rebuilding it | Can we export everything, with its history, in a format we can read without your software? |
| The vendor holds your content and keys | People outside the company may be able to read it | Who can technically read our content and keys, including your staff, and is that a promise or a design choice? |
| The price is unpublished or grows with usage | The bill appears after the team depends on it | What will we pay at three times today's use, and what happens at a limit? |
| A developer memory API is bought for a team job | The team still needs approvals, access for each person and a record, and may have to build or buy them | Who runs this day to day, and does it work in the AI tools the team already uses without code? |

## A model writes memory with no person approving it

Many memory tools let a model decide what to remember. That suits an app that remembers each user's preferences. In a memory a whole company shares, one wrong fact, such as a superseded price or a reversed decision, is read by every agent on the team until someone notices.

What the products' own pages say:

- Mem0's documentation describes a model that reads messages, pulls out facts and adds them as memories, with no step where a person approves a memory before it is saved.
- Supermemory's memory model decides what to learn, what is important and when to forget.
- Letta's agents edit their own memory; its optional review step in memory consolidation does not ask you for approval.
- Cognee lets a model extract facts; when facts conflict it keeps both and a person can settle the conflict afterwards.
- Zep switches agent writes on or off for each connection, with an account-level kill switch.
- Notion Agent learns context over time, such as your role and priorities; its help page links to a view of those memories and does not describe approving them first.

Mensara: people and their agents propose updates. Some permitted changes save after an automatic check. Changes to governed areas, such as decisions, wait for the owner's approval, and you can require approval elsewhere too. Deleting or renaming a file always waits for an approver. The cost is an owner's time spent on approvals.

## No record of who read or changed what

When an agent gives a customer a wrong answer, you need to know what it read and who last changed that record. Without a record you cannot find the cause, and you cannot show anyone what happened.

- Mem0 keeps a change history for each memory, its hosted platform lists recent add, search and delete events for each project, and it lists audit logs on its Custom plan, which has no published price.
- Zep lists audit logs on its Enterprise plan.
- Notion lists its audit log on the Enterprise plan.
- Glean says every control point is logged to an end-to-end audit trail.

Mensara: the Mind keeps a signed, tamper-evident receipt for each read and change, including refusals. A receipt shows who asked, the kind of action, which areas it touched and the outcome. It stores fingerprints, not the file path or the text; a record's version history shows the wording that changed.

## Access is not checked for each person

A shared memory soon holds things not everyone should see: pay, a customer's contract terms, a board discussion. If an agent can read the whole memory, anyone who can ask that agent can read it too.

- Glean says users only see what they are allowed to see, and that it enforces source-system permissions on every read and write.
- Notion says Notion AI can only use content you give it access to, and that Notion Agent has the same permissions you do.
- Mem0's hosted organizations and projects have two roles, Reader and Owner.
- Supermemory can limit an organization member to certain namespaces on its Scale plan.
- Cognee sets read, write, delete and share permissions on each dataset.
- Zep gives people account- and project-scoped roles, and agents' API keys attribute-based access, on its Enterprise plan.

Mensara: each person's agents work within that person's access, and access is checked by software before any model sees a file.

## You cannot export or read the data without the vendor

If you leave, or the vendor changes its terms, the memory your team built should come with you in a form you can read.

- Notion can export a whole workspace as HTML, Markdown or CSV; pages you cannot access are left out.
- Supermemory lets organization owners or admins export up to 25,000 documents and memories as JSON.
- Mem0's export is a hosted-platform feature; its open-source edition runs on your own infrastructure.
- Cognee exports a COGX archive.
- Mem0, Letta and Cognee are open source under Apache 2.0, so you can run them yourself.

Mensara: saved records and their version history are ordinary files in a Git repository. You can keep a private copy in your own GitHub, and copy the files and keep them without Mensara.

## The vendor holds your content and keys

Hosted memory means someone else's servers hold your company's knowledge and often the keys that protect it. That can be acceptable, but you should know who can reach it.

- Zep lists bring-your-own-key and running in your own cloud on its Enterprise plan.
- Mem0 lists on-premises deployment on its Custom plan.
- Supermemory lists a self-hosted option on its Scale plan.
- Notion lists zero data retention with AI providers on its Enterprise plan, and says customer data is not used to train models by default.

Mensara: keys and credentials live on your Mind's server, not in the mensara.ai account system. The owner keeps a Recovery Kit, the key material needed to verify and recover the Mind, somewhere separate from the file backup, and the Git backup excludes private key material. If Mensara operates the server, that separation is not a promise that the operator cannot access it.

## The price is unpublished or grows with usage

There are two cost risks: a price you cannot see until you talk to sales, and usage charges that grow with how much your agents read and write.

- Glean's pricing page shows no price; it offers a demo.
- Mem0 shows a Free plan with stated usage limits and a Custom plan with no price; past the free limits it bills by usage, with a monthly cap you set.
- Zep's Flex plan is $125 a month for 50,000 credits, then $25 for each further 10,000.
- Cognee Cloud's Standard plan is $0 a month plus $0.10 per million tokens, with a one-time $10 starting credit (about 100 million tokens), plus $5 a month for each extra workspace.
- Notion's Business plan, which includes its full AI features, is $20 per member per month billed yearly.
- Supermemory publishes plans from free to $399 a month, and Letta from free to $20 a seat a month for Teams Pro.

Mensara: public pricing is not yet published. The offered price, billing period, tax and terms are shown at checkout. Model usage in the Mind is a different measure from the account charges, so ask what your agents' usage will cost as well.

## A developer memory API is bought for a team job

Mem0, Letta, Zep, Cognee and Supermemory are mainly built for developers and teams building agents that need a memory. They do that job well. Used as a team's shared memory, none of them describes an owner or another person approving each change before it is saved; the nearest are Supermemory's draft, which the person saving can review, and Cognee Cloud's admin decision before memory moves up to a team. Access for each person and a record of activity vary by product and often sit on enterprise plans, so check what the plan you would buy includes.

At the other end, Glean is an enterprise AI platform that connects to more than 275 work apps, and Notion AI works inside a Notion workspace. Check that the product fits the size of your team and the place your work already happens.

Mensara is built for founders and small teams: the agents they already use, such as Claude, Codex, Cursor and ChatGPT on the web, connect to one shared memory over MCP. It is not a memory API for building your own app.

## Where Mensara fits, and where it does not

Mensara gives an AI-native company the Mind it runs on: one memory its people and their agents share, governed by the company, in files it can copy and keep.

- **Fits** a small team that wants its people's AI tools to work from the same records, with approval for governed areas, a signed record of activity and files it can take with it.
- **Does not fit** if you need a memory layer inside an app you are building, a published price before you sign up, or a vendor that lists SOC 2 or HIPAA compliance today; Mensara's public pages list neither.

## Questions

### What is the biggest risk with shared AI memory?

That a wrong fact is saved without anyone checking it and then repeated by every agent that reads the memory. Ask which changes save without a person and whether you can require approval for the records that matter.

### Is a developer memory API like Mem0 enough for a team?

It is built for a different job: giving an app or agent a memory of its users. A team can use one, but Mem0's documentation describes no step where a person approves a change before it is saved, and its SSO and audit logs are on the Custom plan, which has no published price. Check that approvals, access for each person and a record of activity are on the plan you would buy.

### What should a small company ask a shared AI memory vendor?

Seven questions: which changes save without a person; whether every read and change is recorded; whether answers are limited to what each person may see; whether you can export everything in a readable format; who can technically read your content and keys; what you will pay at three times today's use; and whether it works in the AI tools your team already uses without code.

### How do I get the decision approved?

Run a two-week pilot on one recurring job and bring a one-page memo. See [How to get leadership to approve shared AI memory](https://www.mensara.ai/compare/leadership-approval-for-shared-ai-memory).

## Sources

All competitor facts checked 18 November 2026 on each company's own site; Zep, Cognee and Supermemory facts rechecked 19 November 2026.

- Mem0: [home page](https://mem0.ai/), [adding memories](https://docs.mem0.ai/core-concepts/memory-operations/add), [memory history](https://docs.mem0.ai/api-reference/memory/history-memory), [pricing](https://mem0.ai/pricing), [organizations and projects](https://docs.mem0.ai/api-reference/organizations-projects), [platform vs open source](https://docs.mem0.ai/platform/platform-vs-oss), [GitHub](https://github.com/mem0ai/mem0)
- Letta: [memory](https://docs.letta.com/configuration/memory), [pricing](https://docs.letta.com/letta-code/pricing), [GitHub](https://github.com/letta-ai/letta)
- Zep: [pricing](https://www.getzep.com/pricing), [governance](https://help.getzep.com/governance), [agent memory over MCP](https://www.getzep.com/blog/agent-memory-mcp)
- Cognee: [pricing](https://www.cognee.ai/pricing), [permissions](https://docs.cognee.ai/core-concepts/multi-user-mode/permissions-system/overview), [memory versioning](https://docs.cognee.ai/core-concepts/further-concepts/memory-versioning), [GitHub](https://github.com/topoteretes/cognee)
- Supermemory: [how it works](https://supermemory.ai/docs/concepts/how-it-works.md), [namespaces](https://supermemory.ai/docs/concepts/container-tags.md), [pricing](https://supermemory.ai/pricing), [data export](https://supermemory.ai/changelog/request-a-data-export-from-settings/)
- Notion: [pricing](https://www.notion.com/pricing), [Notion Agent](https://www.notion.com/help/notion-agent), [Notion AI security](https://www.notion.com/help/notion-ai-security-practices), [export](https://www.notion.com/help/export-your-content)
- Glean: [pricing](https://www.glean.com/pricing), [security](https://www.glean.com/security)
- Mensara: [About](https://www.mensara.ai/about), [Changes, approvals and receipts](https://www.mensara.ai/how-it-works/changes-approvals-receipts), [Backup, restore and ownership](https://www.mensara.ai/how-it-works/backup-restore-ownership), [Ownership and data flow](https://www.mensara.ai/how-it-works/ownership-data-flow), [Usage, budgets and costs](https://www.mensara.ai/how-it-works/usage-and-costs)

## Related

- [How to get leadership to approve shared AI memory](https://www.mensara.ai/compare/leadership-approval-for-shared-ai-memory)
- [Mem0 alternatives for teams](https://www.mensara.ai/compare/mem0-alternatives)
- [Mensara vs Mem0](https://www.mensara.ai/compare/mensara-vs-mem0)
- [All comparisons](https://www.mensara.ai/compare)

[Sign up](https://www.mensara.ai/auth/signup)
