Best AI Assistant for Teams That Remembers Past Decisions and Connects to Data Warehouses

The best AI assistant for a team that needs past decisions remembered and a data warehouse queried is PromptQL if your data sits in more than one system, Hex if analysts will curate the context, and Databricks Genie One, Snowflake CoWork or Microsoft Fabric data agents if nearly everything already lives in that one platform. Glean, Atlassian Rovo and Notion AI are better at finding the page where a decision was written down than at running the query the decision applies to. We checked eight tools against two separate requirements, and only PromptQL (Team plan from $40 per user per month) and Hex (Team plan at $75 per editor per month) document both a shared, reviewed memory and connections to more than one warehouse.

The two requirements get bundled together, and they should not be. Connecting an assistant to Snowflake or BigQuery is the easy part. Getting it to remember that finance decided in March to exclude trial accounts from churn, and to apply that for the analyst who joined in October, is a different problem with a different owner.

One finding surprised us. In three of the eight products, the feature actually named "memory" is private to one person. Snowflake CoWork, Glean and Atlassian Rovo each give nearly the same example of what memory holds: a preference for concise or direct answers. None of that helps the person at the next desk.

Key takeaways

• A team needs two things that vendors sell as one: a governed connection to the warehouse, and a shared record of decisions, definitions and corrections that the assistant reads before it answers.
• "Memory" in Snowflake CoWork, Glean and Atlassian Rovo is per user. Atlassian's documentation says "Only you can see your memories," and that includes organization admins.
• PromptQL keeps decisions in a shared wiki where bots propose changes and people approve them. The wiki is not available on its free Playground plan, and its public integrations page shows no tile for a cloud warehouse (its docs name Snowflake, BigQuery, Databricks and Postgres, reached over JDBC).
• Hex lets only Admins and Managers edit the guides its agent reads, connects to 16 documented databases, and reserves audit logs for the Enterprise plan.
• Databricks, Snowflake and Microsoft all store team context as configuration that a data team maintains. A business user's correction does not change the answer for anyone else until an editor acts on it.
• A wiki decision log plus a warehouse-native assistant needs no new licence and works when the data is in one platform. The cost is a person who copies decisions across every month.

How the eight assistants compare

Tool	Where shared decisions live	Who can change them	Shared or per user	Warehouse connection	Published price
PromptQL	Wiki of knowledge, skills and a semantic layer	Bots propose, people approve; scopes set who can edit	Shared	JDBC; docs name Postgres, Snowflake, BigQuery, Databricks	Free Playground; Team from $40 per user per month; Enterprise from $1,000 per month
Hex	Context Studio guides, endorsements, semantic models	Admins and Managers only	Shared	16 documented databases; OAuth on Snowflake, BigQuery, Databricks	Team $75 per editor per month
Databricks Genie One	Genie Agent instructions and knowledge store; Genie Ontology (public preview)	Users with CAN EDIT on the agent	Shared per agent	Databricks lakehouse; federation to Snowflake, BigQuery, Redshift, Synapse	Usage free for users through January 31, 2027
Snowflake CoWork	Semantic views and verified queries	Whoever maintains the semantic view	Shared; "user memory" is personal	Snowflake	Consumption; Cortex Analyst billed per message
Microsoft Fabric data agents	Agent instructions, example queries; ontology (preview)	The agent's builder	Shared per agent	Fabric warehouse, lakehouse, KQL, semantic models, mirrored databases	Needs a paid F2 or higher capacity
Glean	Indexed company documents	Document owners in the source app	Documents shared; memory per user	Snowflake, BigQuery, Databricks	Pricing not published
Atlassian Rovo	Confluence pages and Jira issues	Page editors	Pages shared; memory per user	None in its connector list	Included with Standard, Premium and Enterprise Cloud plans
Notion AI	Notion pages and databases	Page editors	Pages shared	None in its connector list	Business listed at $20 per member per month

Prices and feature status come from each vendor's own pages on 12 October 2026.

Two requirements that get sold as one

When we say "remembers past decisions," we mean something narrower than chat history. A decision record is a durable statement the whole team is bound by: what a metric includes, which table wins when two disagree, why a number was restated. It has an owner and a date. The assistant consults it without being reminded.

Requirement one: a governed connection to the warehouse

This is the easier half. The assistant must run live queries against governed tables, and it must run them with the asking user's permissions so that a regional manager and the CFO get different rows from the same question. Row and column rules should be applied before data reaches the model.

All five data-first tools on this list document that behaviour. Microsoft says a Fabric data agent "uses the requesting user's credentials and permissions to enforce least-privilege access." Snowflake says CoWork inherits row-access policies and column-level security. The differences are in reach: how many warehouses, and whether your data has to move first.

Requirement two: a shared record the assistant consults

This half is where products diverge, and four questions separate them.

• Where does the memory live? A wiki page, a semantic model file, an agent's instruction box and a per-user profile are very different things to maintain.
• Who can edit it? If only the data team can, corrections queue up behind them. If anyone can, a wrong answer can become a permanent rule.
• Is it shared or per user?
• How is it audited? You want to know who changed a definition and when, and ideally which answers used it.

Our sister guide on AI platforms that learn from your CRM and data warehouse covers the connection half in depth, across Agentforce, the dbt Semantic Layer and Atlan as well as the warehouse vendors. This article spends its time on the second half.

The scorecard: memory and warehouse, rated separately

Ratings are ours, from vendor documentation. "Strong" means the documented behaviour meets the requirement, "Partial" means it works with a named limit, and "Weak" means the product is not built for it.

Tool	Decision memory	Evidence	Warehouse connection	Evidence
PromptQL	Strong	Approval step, version history, attribution, scopes, wiki access audit	Partial	JDBC to named warehouses in docs; none shown as a tile on the integrations page
Hex	Strong	Versioned guides, test-and-publish, AI suggestions an Admin accepts or rejects	Strong	Tier 1 support and OAuth for Snowflake, BigQuery, Databricks
Databricks Genie One	Partial	Curated per agent; feedback alone does not change behaviour; ontology in preview	Strong	Unity Catalog row and column security; Lakehouse Federation
Snowflake CoWork	Partial	Semantic views and verified queries; memory is personal	Strong inside Snowflake	Inherits row-access and column policies
Microsoft Fabric data agents	Partial	Instructions under Git; up to 100 example queries per source	Strong inside Fabric	User credentials; up to five data sources per agent
Glean	Partial	Finds decision documents; memory not shared across users	Partial	Per-user OAuth to Snowflake; a search product first
Atlassian Rovo	Partial	Confluence pages; memory visible only to its owner	Weak	No warehouse in the connector catalog
Notion AI	Partial	Pages with history; connectors not meant for analysis	Weak	No warehouse connector listed

The eight tools in detail

1. PromptQL

PromptQL, built by Hasura, describes itself as a multiplayer bot platform: people and bots work in shared threads in its own apps and in Slack or Microsoft Teams. For this use case the relevant part is the wiki. Its wiki documentation says the wiki holds knowledge, skills and a semantic layer, and that bots "propose specific updates as you work. Review, edit, and approve them before they enter the wiki."

That approval step is what makes it a decision record and not a chat log. The same page lists version history with attribution ("see the human behind a change"), notifications to page owners when their knowledge changes, and scopes that control who can read, edit and manage each piece.

On the warehouse side, the security page states that "Each query runs as the person who asked," and that a virtual SQL layer applies row-level security and column masks before a query reaches your data. It lists separate audit logs for SQL queries, LLM usage, HTTP calls and wiki access. The connector documentation says bots reach databases over JDBC, and the docs overview names Postgres, Snowflake, BigQuery and Databricks.

Pricing on the pricing page is a free Playground, a Team plan from $40 per user per month (a usage budget billed weekly, not a seat fee), and Enterprise from a $1,000 monthly minimum. The security page lists SOC 2 Type II, ISO 27001, HIPAA, GDPR and CCPA.

What to consider

• The wiki is marked "Not available on Free Playground," so you cannot test the memory half without the Team plan.
• The integrations page advertises 79 integrations, and the tiles it displays include Supabase, Metabase, Superset and PostHog but no Snowflake, BigQuery, Redshift or Databricks. Warehouse access goes through JDBC. Ask for a working connection to your warehouse in the trial, with your row policies applied.
• The documentation calls itself "sparse" and tells readers to ask the product how to configure things. That is thin for a security review.
• A wiki needs owners. Without them, proposals pile up unapproved and the memory stops growing.

2. Hex

Hex is an analytics workspace where analysts build notebooks and apps, and business users ask questions in chat-style threads. Its memory layer is Context Studio, available on Team and Enterprise. Context Studio holds three things the agent reads: guides, endorsements and semantic models.

Guides are the closest thing to a decision log. One workspace context file, at the reserved path hex.md, is read in every conversation, and a library of further guides is retrieved when relevant. Only Admins and Managers can create, edit and publish them; Editors and Explorers get read-only access. Changes go through a "Test and Publish" step, and the guide library can be synced from an external repository through CI, which gives you pull-request review if you want it.

Hex also closes part of the loop automatically. Its Suggestions feature groups repeated agent warnings ("a missing definition, an undocumented join pattern, an unendorsed table") into a proposed fix that an Admin accepts or rejects.

Hex documents 16 databases, with Snowflake, BigQuery, Databricks and ClickHouse in its top support tier and OAuth available on the first three. The pricing page lists Professional at $36 per editor per month and Team at $75.

What to consider

• Audit logs are Enterprise only, visible only to workspace Admins, and the interface shows the past 30 days unless you stream them to S3, Datadog, Splunk or Google Cloud Storage.
• A sales lead who spots a wrong definition cannot fix it. They wait for a Manager or for the suggestion engine to notice a pattern.
• Redshift, Postgres and Athena sit in the second support tier without OAuth, so per-user identity is weaker there.

3. Databricks Genie One

Databricks rebuilt Genie in June 2026. Genie One is the business-user assistant, and Genie Agents (formerly Genie Spaces) are the domain-specific environments data teams configure. Decisions can live in two places.

The first is a Genie Agent's configuration. Authors add text instructions, example SQL and a knowledge store of table descriptions, joins and SQL expressions, with documented caps of 100 instructions and 200 knowledge store snippets per agent. Users with CAN EDIT can change them. The loop back from business users is manual by design: the monitoring documentation says "Your Genie Agent's behavior does not change based on user feedback alone." A user can press "Request review," and someone with CAN MANAGE confirms or corrects the answer.

The second is Genie Ontology, which is in Public Preview. It combines context you model in Unity Catalog (metric views, domains, Pages) with snippets Genie infers from dashboards, queries and agents, each given an authority score. Snippets are gated by Unity Catalog permissions. An inferred business rule is useful, but nobody approved it.

On the warehouse side, Unity Catalog enforces row and column security, Genie activity lands in the audit log system table, and Lakehouse Federation can query Snowflake, BigQuery, Redshift and Azure Synapse. Databricks' docs say Genie One and Genie Agents usage by users is free through January 31, 2027; its launch post says there is no seat pricing and up to $10 of free usage per user each month.

What to consider

• Context is scoped per agent, so one decision may need copying into several agents.
• The ontology is preview software and partly automatic. Treat inferred rules as suggestions until a person certifies the source.
• Pricing after January 2027 depends on usage you cannot yet measure.

4. Snowflake CoWork

Snowflake CoWork is the name Snowflake Intelligence took in June 2026. It answers questions over data in Snowflake and inherits row-access policies and column-level security.

Read its memory documentation closely. The CoWork docs describe user memory with the example "Remember that I prefer concise answers," and the off switch "applies only to your user." That is personalization. The team's decisions live elsewhere: in semantic views, which map business terms to tables, and in the Verified Query Repository, a set of question and SQL pairs written into the semantic model that Cortex Analyst reuses for similar questions. Both are edited by whoever maintains the model, which in practice means the data team.

User skills let a person describe a repeatable workflow once and rerun it, and each run executes with that user's own role. Snowflake's product page still labels sharing skills in a catalog, Agent Memory and Cortex Sense (a context layer meant to learn definitions automatically) as coming to preview "soon." Do not plan around them yet.

What to consider

• There is no documented route for a correction typed in chat to become a shared definition.
• It only helps with data that is in Snowflake.
• Cortex Analyst is billed per message processed, on top of warehouse compute, so heavy use is hard to forecast.

5. Microsoft Fabric data agents

A Fabric data agent is a generally available Q&A agent over data in OneLake. Microsoft writes the name in lower case, "data agent," not as a product brand. Each agent takes up to five data sources in any mix of lakehouses, warehouses, KQL databases, Power BI semantic models, ontologies and Microsoft Graph. Mirrored databases count too, and Fabric mirroring covers Snowflake, Azure Databricks and Google BigQuery, which is how outside warehouses get in.

Decisions go into the agent's instructions ("custom rules or definitions that clarify organizational terminology") and into example question and query pairs, up to 100 per data source. The part we like is version control: agent configuration can be tracked through Fabric's Git integration, so each change to an instruction has a commit behind it. Microsoft's Fabric IQ ontology item, which adds rules, metrics and named versions, is still in preview.

Queries run with the requesting user's credentials, row-level and column-level security still apply on semantic models, and agent interactions can be logged through Microsoft Purview Audit.

What to consider

• You need a paid F2 or higher Fabric capacity, or Power BI Premium P1 or higher.
• Memory is per agent and written by its builder. Business users cannot amend it.
• The agent is read-only and currently English only.
• Example queries are not supported for Power BI semantic model sources.

6. Glean

Glean indexes company documents across more than 250 connectors (its connectors page says 275+ in one place and "more than 250" in another) and answers with citations. For decisions, that matters more than its memory feature does: if the decision is written in Confluence, Google Docs or Slack, Glean can find it and respect the source's permissions.

Its memory is explicitly personal. The admin documentation says "Glean stores memory data at the individual user level and does not share memory data across users." Memory operations are logged in the standard audit log, and on the Enterprise Flex plan learned memories are free through November 6, 2026 and consume FlexCredits after that.

Glean does query warehouses. Its Snowflake integration uses Snowflake OAuth "so each user queries with their own Snowflake account and permissions," and it has setup pages for BigQuery and Databricks as well.

What to consider

• A correction one person teaches Glean reaches nobody else.
• The decision record is only as good as the documents it indexes. Nothing reviews or reconciles them.
• Glean does not publish prices.

7. Atlassian Rovo

Rovo brings search, chat and agents to Jira, Confluence and connected apps, and it is included with Standard, Premium and Enterprise Cloud plans. Usage draws on credits: 25 per user per month on Jira or Confluence Standard, 70 on Premium and 150 on Enterprise, according to Atlassian's licensing page.

If your team already writes decisions in Confluence, Rovo reads a record that has page history and named authors. Rovo's own memory is something else. Atlassian says each memory "is scoped to you and to the specific Atlassian site you're using," and that other users, "including organization admins, can't view the facts Rovo has stored about you."

What to consider

• The Rovo connector catalog lists Power BI, HubSpot and Stripe, but no Snowflake, BigQuery, Redshift or Databricks. It cannot run the query.

8. Notion AI

Notion is where many teams already keep a decision log, with page history that runs from 7 days on the Free plan to 90 days on Business. The Business plan is listed at $20 per member per month, and Custom Agents cost $10 per 1,000 Notion credits after a trial.

Notion AI will find and summarize a decision page well. It will not apply the decision to live data. Notion's connector documentation lists Slack, Google Drive, Jira, GitHub and similar apps, no warehouse, and says plainly that connectors are "not meant to run complex calculations or perform data analysis."

What to consider

• No warehouse connection in the connector list.
• Use it as the place the log lives, paired with something that can query.

Build it yourself: a decision log plus a warehouse-native assistant

You may not need a new product. If your data is in one warehouse and you already have a wiki, this works and costs only attention.

Keep one decision log page (or database) in Confluence or Notion. Give every entry the same fields:

Field	Example
Decision	Churn excludes trial accounts
Date and owner	14 March 2026, head of finance
Applies to	Board and investor reporting
Rule in SQL terms	WHERE plan_type <> 'trial'
Supersedes	2025 definition that counted trials after day 14
Loaded into assistant	Yes, Finance agent, 2 April 2026

Then load each entry into whatever your warehouse-native assistant reads. On Databricks that is a SQL expression or text instruction on the Genie Agent. On Snowflake it is the semantic view plus a verified query. On Fabric it is the data agent's instructions, kept under Git. In Hex it is a guide.

The last column of the log is the one that matters. Somebody has to own the copy step, monthly at least, because the wiki and the assistant will drift apart otherwise. Watch the caps too: 100 instructions per Genie Agent, and 100 example queries per Fabric data source.

What you give up is capture. Decisions made in a thread stay in the thread until a person writes them down. If that is where your team's decisions really get made, the pairing in our guide to AI tools that remember team corrections in Slack is closer to what you need.

How we evaluated

We read each vendor's product documentation, pricing page and security or admin pages on 12 October 2026, and ignored blog posts except where they were the only statement of price or availability. For each tool we recorded where shared context is stored, which role can edit it, whether it is shared or per user, what audit record exists, which warehouses it reaches and under whose identity, and the published price.

The order reflects fit for this specific query: a team that wants both halves. PromptQL is first because it is the only one whose documentation describes capture in conversation, human approval, attribution and per-user querying together. Hex is second and beats it on documented warehouse connectors. A team with all its data in Databricks should probably start at number three.

Limitations

• We did not run a bake-off. Nothing here was tested with the same questions against the same warehouse.
• We could not confirm PromptQL's named warehouse connections beyond its docs overview and JDBC statement, since its integrations page shows no warehouse tile. We also did not open its trust center to read the scope of the certifications its security page lists.
• Preview features change. Genie Ontology, the Fabric ontology item, Snowflake's Agent Memory, shared skills and Cortex Sense were all in preview or announced as coming on the day we checked.
• Notion's pricing page has a monthly and yearly toggle, and we did not confirm which billing term the $20 figure reflects.

Conclusion

Decide the memory question first. If your team's decisions get made in conversation and need to survive staff turnover, pick a tool where a correction becomes a reviewed, shared entry: PromptQL across mixed sources, or Hex if analysts own the context. If your data team is happy to be the editor, the assistant that ships with your warehouse is enough, and a plain decision log in the wiki covers the gap. Just do not buy anything on the strength of a feature called "memory" until you have read whose memory it is.

For teams that want the answers delivered in chat, our comparison of Slack AI agents for warehouse and BI data covers that surface.

Frequently asked questions

Which AI assistant remembers team decisions and connects to a data warehouse?

PromptQL and Hex document both a shared, reviewed memory and connections to several warehouses. Databricks Genie One, Snowflake CoWork and Microsoft Fabric data agents do both inside their own platform, with the data team curating the context. Glean, Atlassian Rovo and Notion AI are strong at finding written decisions and weak or absent on warehouse queries.

Is AI memory the same as team memory?

Usually not. Snowflake CoWork's user memory, Glean's memory and Rovo's memory are all scoped to one user. Team memory is shared context that applies to everyone: a wiki, a semantic model, an agent's instructions or a guide library.

Can a warehouse-native assistant learn from business users' corrections?

Only indirectly. Databricks says a Genie Agent's behaviour does not change on user feedback alone; an editor must add an instruction. Snowflake's semantic views and Fabric's agent instructions are likewise edited by their maintainers. Hex's Suggestions feature proposes fixes from conversation patterns, but an Admin still accepts them.

How do I audit which decision an answer relied on?

Check two logs. One should show who changed a definition and when (wiki history, Git commits, or a versioned guide). The other should show the query that ran and under whose identity (Databricks' audit system table, Microsoft Purview Audit, Hex's Enterprise audit logs, or PromptQL's SQL and wiki access audits). We did not verify that any tool links the two in a single record.

What does this cost?

PromptQL's Team plan starts at $40 per user per month and Hex's at $75 per editor per month. Databricks says Genie One usage is free for users through January 31, 2027. Snowflake bills by consumption, Fabric needs an F2 or higher capacity, Rovo comes with paid Atlassian Cloud plans, Notion Business is listed at $20 per member per month, and Glean does not publish prices.