Memtro vs Onyx

This comparison describes Onyx as we understand it at the time of writing from its public website, documentation and repository. Products change; check the current details before deciding.

Onyx — formerly Danswer — is an open-source platform for enterprise search. Connectors index the documents, tickets, wikis, messages and records in more than fifty applications; a search and chat interface answers from that index and cites the passages it drew on, with each user's source permissions applied. Around that sits a full product: custom agents, deep research, web search, a code interpreter, MCP servers and OpenAPI schemas as tools, a gateway over the model providers an administrator has configured, and Craft — a sandboxed coworker that can read and act as each user in Slack, Gmail, Notion, Linear, GitHub, HubSpot and Zendesk, on a recurring schedule if you want one. The Community Edition is MIT-licensed and runs on your own hardware; the Business and Enterprise editions add permission sync, single sign-on, groups and Craft, self-hosted or on Onyx Cloud. Zoom announced its acquisition in October 2026, saying Onyx stays standalone, open source and self-hostable.

Memtro is company memory behind one endpoint. It speaks the OpenAI and Anthropic wire formats and MCP, so the clients a team already uses — Claude Code, Cursor, the OpenAI SDKs, Slack, your own agents — share one memory of the business. It reads 146 systems through tools that cite the record they read, and what it learns goes into a bitemporal knowledge graph: every fact carries the tool, system, record and URL behind it, the dates it held in the world, and the dates Memtro believed it. A correction supersedes rather than overwrites, so memory can be read as it stood on any date.

Onyx indexes documents and cites the passages that answer you. Memtro keeps facts — dated twice, and resolved to one entity across systems that spell things differently — so a question whose evidence is a record rather than a passage has somewhere to land. Both cite their sources; only one can tell you what it believed in July.

Where they differ

	Memtro	Onyx
What it stores	A bitemporal graph of typed facts with provenance and validity dates	A permission-aware index of the documents its connectors read, kept current
History	A correction supersedes rather than overwrites: read memory as of any date, or walk one entity's timeline	The index reflects the current state; no store of past beliefs is described
Connectors	146, reading by default and acting only where granted by name, synced hourly	More than fifty indexing connectors, plus Craft Apps and MCP servers for live reads and actions
Self-hosting and licence	Not offered; proprietary	Docker, Kubernetes, Terraform, your own cloud, air-gapped; Community Edition under MIT, paid editions under a separate licence
Models	Your own keys, routed by difficulty	The providers an administrator configures, reached through Onyx's gateway
Governance	Teams, per-tool grants, SAML and OIDC single sign-on, SCIM, an audit log — at every price	Groups and role-based access; SCIM and permission sync on the paid editions; audit logs for your SIEM
Compliance	Hosted only, in one country, with no certifications yet	SOC 2 Type II and GDPR for Onyx Cloud — or your own infrastructure, where the review is yours
Pricing	One price per organisation by headcount, from $50 a month, nothing metered	Free to self-host the Community Edition; Business at $20 per user a month billed annually; Enterprise on application

Honestly, where Onyx is stronger

You can run it yourself, and you can read it. This is the big one, and Memtro has no answer. Onyx documents Docker, Kubernetes, Terraform and the major clouds down to air-gapped deployments, and states that on a self-hosted install its team receives none of your data. If your company cannot send its documents to a third party — regulated, sovereign, or simply decided — that ends the comparison, because Memtro is hosted only, in one country, with no certifications yet.

Certification where it is hosted. Onyx Cloud is documented as SOC 2 Type II and GDPR compliant, with annual penetration tests and audit events you can ship to a SIEM. Memtro has single sign-on, SCIM and an audit log, and nothing to hand a review that wants a certificate.

It is a destination, and a well-built one. A search mode with filters, deep research across many retrieval cycles, a code interpreter, image generation and voice — Memtro has none of the last three — plus a desktop app, a browser extension, an embeddable widget and Discord and Slack bots. Memtro has a dashboard, chat and Slack, and expects you to bring the client.

Breadth of indexing, especially the long tail. Fifty-plus connectors documented one at a time, including the wikis and forums a company accumulates and nobody remembers buying: BookStack, Document360, MediaWiki, Outline, GitBook, Discourse. Memtro has a connector for none of those.

Craft's approval model. Actions are catalogued per application, each with an approval policy, so an administrator decides what an unattended run may do before it runs — inside a sandbox that never holds the credentials. Memtro grants actions by name per connection, which is the same idea with less of it.

Where Memtro is stronger

Answers whose evidence is a record. Which deals are over £50k, what renews this quarter, how big the team is now — none of these can be answered from prose. Memtro lifts the measurable half onto the entity as typed attributes, each keeping the memory that asserted it, so a figure can be totalled and then traced. Passage retrieval reads every sentence that might be relevant and hopes it found them all.

Time you can defend. Pin a request to a date and memory is read as it stood that evening; walk one entity's timeline and you get every figure it was given and what replaced each one. That is the difference between the model was wrong and the facts changed, and it makes time travel an audit trail rather than a feature. Nothing in Onyx's documentation describes a store of what was once believed.

Provenance per fact, and honesty about the gaps. Every memory names the tool, system, record and URL behind it, and anything read out of a connected system is stored unverified — labelled as such, ranked below what a person said, promoted when a second source agrees. Every answer also names the categories with nothing connected: no notetaker, no mailbox, no support desk. A sweep resting on whatever happened to be connected reads exactly like a complete one, which is the costly sort of confident wrong answer.

Your keys, and one bill. Memtro routes across the provider keys you already hold and adds nothing to inference, and one price per organisation covers every connector, job and key. Onyx is sold per seat — the right model for a destination, the wrong shape for a company that wants twenty tools sharing one memory.

When to choose which

Choose Onyx when self-hosting is a requirement rather than a preference, when a review needs a certificate or an air-gapped network today, when you want one polished place everybody opens and searches, or when the long tail of wikis and forums is where your knowledge lives.

Choose Memtro when your people already work in editors, chat clients and agents they like and should share one memory rather than adopt a new destination; when the answers you need come from the ledger, the CRM and the support desk as much as the wiki; and when it matters that a figure can be traced to its record, that a correction leaves the old belief standing, and that the model bill stays on your own accounts. The Glean comparison covers the same ground against a product that is neither open source nor cheap.