Memtro vs Merge
A unified API, MCP tool platform, model gateway and employee governance from an integrations vendor, versus company memory with registers, routing on your own keys and a flat price.
This comparison describes Merge as we understand it at the time of writing from its public website and documentation. Products change; check the current details before deciding.
Merge began as a unified API: one set of common data models over HR, payroll, ATS, accounting, CRM, ticketing and file-storage systems, with an embedded auth component ("Merge Link") so a software product can let its customers connect their own accounts, plus syncing, webhooks and field mapping. It now sells three more things under the banner "the universal context engine for agents": Agent Handler, a remote MCP server over "thousands of pre-built tools" with built-in auth, tool bundles scoped per agent, audit logs and data-loss-prevention scanning; Gateway, an OpenAI-compatible endpoint across the major providers with routing by cost, latency or quality, failover and budget caps; and Merge for Workforce, which pushes approved models, MCP connectors and "skills" to employees' AI tools through an IT-managed endpoint with identity-provider provisioning.
Memtro is company memory behind one endpoint. It speaks the OpenAI and Anthropic wire formats and MCP, routes each request by difficulty across the model keys you already hold, and reads 144 systems through tools that cite the record they read. What it learns is kept in a knowledge graph with provenance and a timeline, and from that graph it keeps registers nothing else keeps: the commitments people made, the contradictions between what the company believes, the jargon it uses, the decisions behind its code, and the figures it measures. It is priced per organisation, by headcount, with nothing metered.
Where they differ
| Memtro | Merge | |
|---|---|---|
| Shape | One endpoint (OpenAI, Anthropic, MCP) in front of your models and systems, with memory | Four products: a unified data API, an MCP tool platform, a model gateway, and employee governance |
| Who it is for | A company that wants the AI tools it already uses to share one memory of the business, and builders of agents who want that memory and those tools | Software companies shipping integrations to their customers, and enterprises governing employee AI; agent builders through Agent Handler |
| Memory | A bitemporal knowledge graph: facts with sources and validity dates, corrections that supersede, readable as of any date | A "live graph of your people, customers, tickets and documents" kept current from connected systems; no provenance per fact or history described |
| Registers | Commitments, contradictions, jargon, decisions, metrics, each a view of the graph | None |
| Connectors | 144, reading by default and acting where an action is granted by name, with hourly sync into memory | Hundreds of integrations in the unified API; 300+ MCP connectors; a public MCP server can be imported |
| Unified data models | No: each system keeps its own shape, and the graph resolves the same person or company across systems | Yes: common models per category, field mapping, passthrough for the rest |
| Embedded auth for your customers | A connection-request flow inside an organisation; no embeddable component yet | Merge Link, with sandboxes and webhooks on sync |
| Routing | By difficulty: five levels, a ladder per level with fallbacks, a coding ladder, the reason on every response | By cost, latency or quality, a build-your-own router with custom benchmarks, the reason on every response |
| Cost control | Token allowances per level, per organisation and team; a lock that keeps requests on auto | Budget caps by project, team or customer tier; context compression; pass-through billing |
| Governance | Teams, permissions, per-tool grants on shared connections, a request log by user, team and model | SAML SSO, SCIM provisioning, RBAC, audit logs of every tool call, DLP policies that block, redact or mask |
| Compliance | Hosted in the United Kingdom; no certifications yet | SOC 2 Type II, ISO 27001, HIPAA, GDPR; US, EU and APAC regions; single-tenant and self-hosted options |
| Scheduled work | Jobs: prompts the agent runs on a schedule with its owner's tools, reporting to a chat or a channel | Not offered |
| Chat clients | Dashboard chat, Slack, and anything that speaks the OpenAI or Anthropic API or MCP | Any MCP-compatible tool through Agent Handler and Workforce |
| Pricing | Free for 7 days, then one price per organisation by headcount: connectors, jobs, memory and keys never metered | Unified API from $650 a month for 10 linked accounts and $65 each after; Gateway at a 5% markup on inference or $0.05 per million tokens on your own keys; Agent Handler on credits per tool call |
Honestly, where Merge is stronger
Scale and certification. Merge has been building integrations for years, holds SOC 2 Type II, ISO 27001 and HIPAA, offers EU and APAC regions, single-tenant deployments and a self-hosted option, and has the governance surface a large enterprise's security review asks for: SAML, SCIM, role-based access, DLP on every tool call, and an audit log of who called what with which arguments. Memtro has none of the certifications and is hosted in one country.
If you are a software company that needs to ship HR, ATS or accounting integrations to many customers, Merge's unified models and Merge Link are the point: one data shape per category, an auth flow your customers complete inside your product, and webhooks when their data changes. Memtro does not try to be that.
Merge's gateway also lets you define what "best" means with your own benchmarks and weights, and compresses context to cut token cost, both of which Memtro lacks.
Where Memtro is stronger
Memtro starts from a different question. Merge gives an agent tools; Memtro gives it a memory of the company that the tools feed, and it keeps that memory honest: every fact carries its source, a correction supersedes rather than overwrites, and the graph can be read as it stood on any date, so you can show what the assistant knew when it acted. Merge's context layer is described as a live graph of records; nothing on its site describes provenance per fact or history.
From that memory come the registers. A commitment made in a meeting is recorded, checked against the other systems, and settled with evidence or sent to a tracker. Two beliefs that cannot both be true are found overnight and put to a person. The words the company uses its own way are defined once and ride in every answer. Decisions are read from pull requests and kept against the code. A figure worth measuring becomes a metric on a schedule. None of these exist in Merge.
And the economics differ. Memtro routes across the keys you already hold and adds nothing to inference; one price per organisation covers every connector, job and key. Merge meters linked accounts, marks up tokens on its gateway, and charges per tool call on Agent Handler, which is the right model for a vendor and an expensive one for a team whose usage grows.
When to choose which
Choose Merge when you are shipping product integrations to customers and need unified models, embedded auth and webhooks; when an enterprise security review demands certifications, regions and DLP today; or when you want employee AI governance with identity-provider provisioning.
Choose Memtro when a team's existing tools should share one memory of the company, with every answer cited to a record; when promises, contradictions, decisions and figures should be kept as well as data; and when model spend should stay on your own accounts at one flat price.