Governed memory with citations
Cortex records what your team decided and on what evidence, and answers with citations from rooms the asker was in. Apache-2.0, pre-1.0.
Cortex keeps a record of each decision
It reads meeting transcripts and documents, extracts each decision with the quote it rests on, and links a reversal to the decision it replaces.
- Ingest from a drop folder, or Markdown, Obsidian, PDF, Word, HTML and text exports
- A bi-temporal decision graph: when a decision held, and when Cortex learned it
- A superseded decision gets an end date and an audited link, and stays in the history
Cortex’s hybrid retrieval and confidence gate
Each question runs two searches over what the asker may see, fuses them by rank, and must clear a gate before an answer is written.
Answer from 0.67, suggest from 0.60, silence below: precision over coverage. On the keyless hashing embedder the gate never fires, and Cortex declines.
Every Cortex answer cites its sources
Above the gate, the model answers from the retrieved text with inline citations, and is told to say so when the sources do not hold the answer.
At ingest, a fabrication check rejects any decision whose quote is not word for word in its source.
Conscience checks every Cortex answer
Conscience Tier 1 is default-deny per asker. It checks every candidate before ranking, so a room the asker was not in looks the same as nothing there.
- One law, compiled to SQL for Postgres and to an in-memory predicate for the cache
- A plug-in backend’s hits are checked again before serving
- An append-only audit of what was admitted and what was not
Cortex’s refusal logic will be open
Conscience Tier 1 will ship in the Apache-2.0 engine, so anyone will be able to read what a node declines to show. The open-source release is being prepared.
- To be open, Apache-2.0The engine, Conscience Tier 1, the MCP server, the CLI, the simulation harness and the drop folder
- CommercialVendor integrations such as Google, Slack and Notion, the Slack bot, the Cortex Launcher and the SDK
Once the code is public, outside pull requests will wait for the contributor licence agreement, which is still a draft.
Where Cortex runs
Storage, retrieval and governance run on your hardware. Inference runs there too, or at a hosted provider on your own key.
- In appliance mode, Cortex refuses to start with a model endpoint outside your network
- The local model plane has not yet run end to end
- There is no managed Cortex service
Cortex’s MCP server, CLI and HTTP API
Agents use five MCP tools, people the cortex command, and services the node API. Every request is answered as one asker.
cortex init prints the line that adds the server to Claude Code. Over HTTP, each request carries a per-seat token.
How Cortex compares with Mem0, Zep and Letta
A design comparison from each system’s published documentation, as of July 2026. It is not a measured benchmark.
Measured retrieval results are below.
| System | How it stores memory | When a fact is contradicted | History | Source |
|---|---|---|---|---|
| Cortex | A decision graph indexed over the source text, which is kept as the record | An audited supersedes link: who changed which decision, when and why, with evidence | Kept | Cortex research note |
| Mem0 | Atomic facts, with an optional graph | Overwrite or delete (add, update and delete operations) | Lost | Mem0 research |
| Zep / Graphiti | A temporal knowledge graph and episodes | Bi-temporal edge expiry, after an invalidation the model judges at ingest | Kept | Zep paper |
| Letta | Core memory blocks, a recall log and archival memory | The agent rewrites its memory blocks | Partial | Letta blog |
Retrieval against standard baselines
Five retrieval methods on Cortex's own golden set: 57 questions over a 14-document synthetic corpus, run in November 2026. Every method used the same embedder, the same access rules and the same user.
| Method | recall@5 (46 answerable) | MRR@5 | Stale document ranked above current (of 5) | No-answer questions given passages (of 11) |
|---|---|---|---|---|
| BM25 (keyword) | 0.978 | 0.951 | 0 | 11 |
| Vector top-k | 0.978 | 0.871 | 0 | 11 |
| Vector + rerank | 1.000 | 1.000 | 0 | 11 |
| Cortex hybrid, no rerank | 1.000 | 0.928 | 0 | 11 |
| Cortex as shipped | 1.000 | 1.000 | 0 | 11 |
Abstention
The baselines have no way to abstain, so they return passages for every question, including the 11 with no answer in the corpus. Cortex runs a confidence gate on top. On those 11 it answered 5, gave a hedged suggestion on 5 and abstained on 1. It also abstained wrongly on 3 of the 46 answerable questions (2 of 46 without the reranker).
What this shows
Cortex is competitive on retrieval, and abstention is where the work is. With 14 documents, a top-5 cut covers more than a third of the corpus, so every method is near perfect and one question moves recall@5 by 2.2 points. Cortex ties vector search with a reranker: this set shows no retrieval advantage. The gate is what the baselines lack, and today it still lets 5 of 11 no-answer questions through as answers.
Disclosures
- The Cortex team wrote both the questions and the corpus. This is not an independent benchmark.
- The set is small: 46 answerable questions, 11 with no answer, 5 that check a stale document never outranks the current one.
- One embedder (nomic-embed-text-v1.5, 256 dimensions), one reranker (Qwen3-Reranker-8B, hosted on Fireworks) and one machine.
- BM25 is textbook Okapi (k1 = 1.2, b = 0.75), in memory, with no stemming.
- Cortex's retrieval settings and gate thresholds were tuned earlier against evals that overlap this corpus.
- The metrics were the same on two full runs. Across repeated calls, the hosted reranker reordered results for 10 of the 57 questions.
- Latency is left out: it was measured on one development machine with hosted API round trips and is not a production figure.
- The per-question results will be published with the open-source release.
An independent, larger benchmark (LongMemEval against open-source Mem0 and Zep/Graphiti) is the next step.
How Cortex is tested
Beyond the retrieval results above, a simulation harness tests the claims on this page in thirteen gates, across layers L0 to L7.
- Citation faithfulness: whether the cited text supports the answer