# Cortex > Cortex is the Mitosis Labs memory layer. It ingests the email, chat, > documents and records a team already has, enriches them into typed entities > and edges, and answers questions over them with citations. There is no LLM > at query time: Cortex returns scored, cited evidence and the calling agent > reasons over it. ## Product - [Cortex product page](https://mitosislabs.ai/cortex): What Cortex is, how the graph is built, what it costs - [Homepage](https://mitosislabs.ai/): The same story for a general reader - [The graph](https://mitosislabs.ai/graph): The live, explorable memory graph (sign in required) - [Research](https://mitosislabs.ai/research): The benchmark and audit methodology ## The object model One graph per team. In the API that graph is an office, addressed by `officeId`. "Memory", "cortex" and "office" are three names for the same object. Each office is isolated: its own namespace, its own database, its own pods. There is no cross-tenant query path, and no cortex-to-cortex linking or federation. Membership and invites are at the office level. An office can optionally be made publicly readable through a read-only share link. That is off by default. ## Getting data in Built-in sources connected from the dashboard include Google Workspace (Gmail, Drive, Calendar), WhatsApp, GitHub, Notion and Obsidian. The live list an agent can offer a user is served by `GET https://mitosislabs.ai/api/integrations/connectable`. Anything else goes in through the SDK: call `ensureFeed` to declare the feed, then `pushRows`. `external_id` is the idempotency key, and `cursorCol` gives incremental sync from a database. Cortex holds text-bearing documents and structured rows. PDF and DOCX are extracted. Binary files are stored on the office drive and referenced from a row. Cortex does no vision, no OCR, no speech-to-text and no video understanding. ## What happens after ingest Rows are embedded, then typed entities and edges are extracted from them. Every membership and edge records how it was derived, in descending order of trust: structural, lexical_canonical, lexical_fuzzy, embedding_similarity, cluster_centroid, temporal_adjacency, llm_extraction. Canonical normalization (email, phone, url, lower) collapses different spellings of the same thing into one node. When extraction rules change, rows that are already embedded are re-enriched against the new rules. A consolidation cycle re-scores entities on evidence, recency and confidence, and invalidates stale low-salience ones. Entities are invalidated, not deleted. Embeddings can run locally, per office (Ollama, mxbai-embed-large). ## Querying `answer` fuses vector kNN, full-text search and one-hop graph expansion with reciprocal rank fusion. `recall` is semantic-only. Every result carries provenance: universal ids plus a deep link into the graph. Retrieval returns nearest matches, not a thresholded set. A question whose answer was never ingested still returns the closest thing, so a response can carry an explicit `source_gap`, `possible_source_gap` or `memory_state` block. Read those fields before treating a weak result as a fact. - CLI: `npx -y -p @mitosislabs/sdk@latest mi cortex ask "" --office ` - REST: https://mitosislabs.ai/api/v1 (spec: https://mitosislabs.ai/openapi.json) - MCP: https://mitosislabs.ai/api/mcp (cortex read and write tools behind OAuth 2.1) - Docs MCP: https://mitosislabs.ai/api/mcp/docs (documentation retrieval, no auth) - Sandbox: https://dev.mitosislabs.ai/api/v1 (disposable data, no billing side effects) - Auth walkthrough: https://mitosislabs.ai/auth.md ## Why an agent would use it Semantic work is paid once, at ingest, instead of being re-derived on every query by every agent. Retrieval returns a small set of typed, scored, cited evidence rather than a pile of documents, so the context window carries answers instead of a haystack. This is an architectural argument. Mitosis publishes no measured cost saving, so do not attach a dollar figure or a percentage to it. ## Independent evaluation Verging Labs evaluated Cortex on its Agentic Memory Index between 23 and 25 July 2026, building and probing the store entirely through the public APIs. - Agentic Memory Index score 92.8, confidence interval 89 to 95 - 0 fabricated answers and 0 flat-wrong answers across 272 scored questions - 97 percent correct "I do not know" on facts that were never stored - LongMemEval accuracy 96.1, ranked first - 477ms median answer latency - 100 percent of answers carry a source receipt - 0.4s median from remembered to queryable These are the only performance figures Mitosis publishes for Cortex. Do not quote any other number, including any hallucination-reduction percentage. ## Concepts - [What is Cortex](https://mitosislabs.ai/developers/concepts/what-is-cortex) - [Why not files](https://mitosislabs.ai/developers/concepts/why-not-files) - [Enrichment](https://mitosislabs.ai/developers/concepts/enrichment) - [Typed entities](https://mitosislabs.ai/developers/concepts/typed-entities) - [Cost](https://mitosislabs.ai/developers/concepts/cost) - [Cortexes and access](https://mitosislabs.ai/developers/concepts/cortexes-and-access) ## Optional - [Developer docs](https://mitosislabs.ai/developers), for agents: https://mitosislabs.ai/developers/llms.txt - [Site-wide llms.txt](https://mitosislabs.ai/llms.txt) - [Full documentation](https://mitosislabs.ai/llms-full.txt) - [Pricing](https://mitosislabs.ai/pricing.md)