Name: Cortrix
Author: Cortrix

Start with the workflow

What is Semantic Storage?

Semantic storage connects information with the context an application needs to use it. Cortrix explores this through retrieval, memory records, and links back to sources. Read the guide, or start with the part of your workflow that interests you.

Agent workflows

An agent needs to read source material, use tools, and keep track of what happened. When evaluating agent storage, start with one task and inspect what comes back from a query.

Try the Docker quick start →

Context for a task

A useful answer needs relevant material at the right moment. Context storage keeps that material available; retrieval still has to decide which passages belong in the prompt.

Read the RAG explainer →

Memory across tasks

Some experience is worth carrying into the next task. Memory storage preserves it, but the system still needs to check whether an old observation or lesson applies.

Read the Reflexion explainer →

Cortrix is in active pre-release. Retrieval and memory interfaces have different maturity levels; authentication, authorization, and production readiness are not established. See current capabilities and limits.

Read the research

Five classic papers trace the ideas from linked data to agent memory. Start with DBpedia or browse the series.

Explore the retrieval benchmarks →

From Source Material to Semantic Records

Cortrix turns documents, code, workflow events, and retrieval evidence into semantic records built for agentic systems.

Cortrix architecture overview with source-processing, retrieval, record, storage, and integration paths labeled by evidence boundary

The diagram is a source overview, not a production-readiness or scale claim. Interface, memory, and integration paths retain their published evidence boundaries.

Source-First and Database Paths

Use the documented loopback-only Docker path for a local first-value evaluation. The pgCortrix path is present in source; compatibility and migration guidance still require review.

Feedback Signals for Review

Observability and feedback signals help teams inspect what supported an agent answer. Automatic learning from those signals is not a current capability.

Scoped Connector Paths

MCP, HTTP, SDK, and other connector surfaces are present in source with different review states. Check the pinned status before adopting a path.

How the pieces fit together

Cortrix brings retrieval, memory records, and source context into a shared design. These are the parts to consider alongside the tools you already use.

Agent-Native

Designed around agent data workflows: semantic processing, memory, and traceability are treated as shared storage concerns rather than scattered integration code.

Semantic Storage

Documents are parsed, chunked, embedded, and indexed in a shared semantic layer so agent workflows can query meaning and source context together.

Built to Fit Existing Stacks

Designed to work with existing databases, tools, and agent workflows while giving retrieval, memory, and audit a shared semantic layer.

Explore the building blocks

Cortrix organizes semantic ingestion, hybrid retrieval, memory records, and source context in one pre-release design. The cited commit contains api/openapi.yaml; file presence and identity do not establish runtime, production, or security readiness.

Semantic Processing Chain (SPC)

Parsing, OCR fallback, chunking, enrichment, embedding, and indexing paths are present in source. Parser and production readiness require review at the pinned source.

Review required

Hybrid Query + Reranker

Vector, BM25, RRF fusion, and reranking paths are present in source for retrieval evaluation. Treat runtime behavior as review required at the pinned commit.

Review required

AI Interaction Memory

Memory APIs and typed records are present in source. MEM02 fact extraction and security isolation are Blocked; user fields are logical scope, not authorization.

Mixed status

Agent Observability

Session, trace, and agent-header surfaces are present for inspection and remain under review. Automatic retrieval learning is a future direction, not a current claim.

Review required

Source-Level Traceability

Cortrix is designed for inspectable source context across chunks, documents, and turns. Treat end-to-end citation and attribution behavior as review required.

Review required

Namespaces + Cross-NS Query

Namespaces provide logical query scope for projects and workflows. They are not authentication, authorization, ACL, or tenant-security boundaries.

Logical scope

Retrieval Quality Foundations

Published evidence covers retrieval quality only. Parent-child and broader retrieval patterns remain subject to their pinned source and review status.

Evidence bounded

pgCortrix Extension

pgCortrix is present as a PostgreSQL integration path. A reviewed compatibility matrix and migration guide are not published.

Review required

MCP Server

The MCP server surface is present in source for agent-tool evaluation. Its current compatibility and production readiness require review at the pinned source.

Review required

Local Embedding Path (bge-m3)

ONNX Runtime and BGE-M3 are present as a local embedding path. Verify the pinned configuration and review status before adoption.

Review required

Connector Paths

HTTP, MCP, Python SDK, adapter, import, and PostgreSQL paths are represented in source with different maturity states; no security enforcement is implied.

Mixed status

Python SDK + REST + Web UI

The cited commit contains the declared api/openapi.yaml specification file. The Python SDK, local server, built-in Agent, and dashboard paths are present in source and remain under review.

File identity + review

Start Cortrix locally

You need Git, Docker, Docker Compose, and curl. No .env file, LLM key, host-side model tools, manual model download, model conversion, or separate bootstrap step is required. The first start downloads about 1.17 GB of pinned model assets and can take several minutes; later starts reuse the cached volume.

Agent-assisted local setup

Install Cortrix with your AI Agent

Copy one version-pinned setup brief. Your agent checks the local prerequisites, starts the loopback-only Docker stack, runs a real reranked query, and returns a verification report.

Local only No LLM key required Real embedding + reranking Docker required

Inspect the Setup Contract

For AI Agents with local filesystem and terminal access. The contract does not authorize sudo, system dependency installation, non-loopback exposure, secrets, destructive cleanup, or production deployment.

Read the full prompt before copying

Agent task

Install Cortrix locally under the cortrix-agent-quickstart/v1 contract at:
https://cortrix.ai/agent-quickstart.txt

Use only https://github.com/cortrix/cortrix.git at release v1.0.0-rc.2 and expected commit d298d6b7e64c67551f567e7cddd514bae1adb57b.

Work only in a new local directory. Before changing state, verify Git, Docker, Docker Compose, curl, available disk, required network access, and that 127.0.0.1:8420 is available. Stop and report if preflight fails or the target directory already exists.

Do not use sudo, install system dependencies, request secrets, change firewall or network settings, expose a non-loopback port, disable ONNX, embedding, reranking, readiness or integrity checks, or delete existing Docker resources.

Follow the contract's exact clone, identity, Compose, readiness and reranked-query steps. Leave the verified service running. Return the required PASS or FAIL report, and mark PASS only when every contract assertion succeeds.

Manual path

Run it yourself

Use the same loopback-only Docker path directly in your terminal.

1

Clone and start Cortrix

Terminal

git clone --branch v1.0.0-rc.2 --depth 1 https://github.com/cortrix/cortrix.git
cd cortrix
test "$(git rev-parse HEAD)" = "d298d6b7e64c67551f567e7cddd514bae1adb57b" || exit 1
CORTRIX_SOURCE_REVISION="$(git rev-parse HEAD)" \
  docker compose -f deploy/docker-compose.yml up --build --wait

2

Check readiness

Terminal

curl -fsS http://127.0.0.1:8420/api/v1/system/health/ready

3

Run a reranked query

Terminal

curl -fsS -H 'Content-Type: application/json' \
  -d '{
    "namespaces": ["demo"],
    "query": "What does semantic storage keep close to the agents that need it?",
    "top_k": 5,
    "rerank": true
  }' \
  http://127.0.0.1:8420/api/v1/query

This loopback-only path uses real BGE-M3 embedding and bge-reranker-v2-m3 reranking on CPU. It leaves external LLM roles and the built-in Agent disabled, publishes neither the metrics port nor the Agent port, and does not establish parser coverage, authentication, internet-facing deployment, benchmark quality, or production readiness. A successful query returns source-backed demo content with numeric rerank_score values.

Terminal

# Stop the service; add --volumes to remove cached models and data.
docker compose -f deploy/docker-compose.yml down

Built for Real Agent Data Workflows

Cortrix focuses on the shared data layer behind agent workflows: documents, retrieval context, memory, feedback signals, and traceability that are often stitched together with custom glue code.

Agent Workflow Storage

Agent workflows need persistent semantic memory and inspectable source context. Cortrix gives builders a shared storage foundation for evaluating those capabilities.

Reviewable Agent Context

Source-level traceability helps teams inspect which documents, chunks, and turns contributed to an answer or workflow step.

Workflow Orchestration

Use Cortrix as a shared semantic layer that MCP, REST, Python SDK, and framework-based workflows can call without each agent owning a separate retrieval stack.

Developer Knowledge Workflows

Use Cortrix as a semantic layer around documents, workflow context, and agent memory while keeping existing databases and tools in place.

How Cortrix Fits

A scope-aware view of a shared semantic storage layer versus maintaining many glue-code paths.

Capability	Cortrix	Typical Glue-Code Stack
Document Ingestion	Review required parser and ingestion paths present in source	Often handled through separate parser and chunker components
Embedding	Review required local embedding path present in source	Often handled through an external model or service
Vector Search	Review required vector path present in source	Often delegated to a separate vector database
Keyword Search	Review required keyword path present in source	May require a separate search component
Hybrid Fusion	Review required fusion path present in source	Often implemented as custom orchestration code
Reranker	Review required reranker path present in source	May require an external service or custom step
Cross-Namespace Query	Review required logical cross-namespace query path	Often handled in client-side orchestration
Advanced RAG	Future direction no release commitment	Often handled by framework plugins or custom pipelines
AI Memory	Mixed status typed records present; MEM02 extraction Blocked	Often handled as a separate memory service or custom store
Source-Level Traceability	Review required source-context paths present in source	Scope varies by stack and tracing implementation
Agent Observability	Review required session and trace surfaces present	Often requires separate instrumentation
Retrieval Feedback Learning	Future direction automatic learning is not current	Usually implemented as a custom feedback loop
PostgreSQL Integration	Review required pgCortrix source path	Often runs as a separate service
Workflow Connectors	Mixed status exact OpenAPI file present; interfaces under review	Usually framework-specific
MCP Server	Review required MCP server present in source	Availability depends on the chosen stack
Scope Boundary	Evidence states Verified / review required / Blocked / future	Usually spread across separate docs
Deployment	Loopback-only Docker Quick Start; other deployment paths require review	Often multiple services

Contribute and Follow Cortrix

Use the published issue, contribution, security, and benchmark-challenge routes on the project page, and follow the official update channels.

Project participation Benchmark Methodology Agent-readable Docs