# Cortrix > Cortrix is an active pre-release, local-first, open-source agent-native semantic storage project for agent builders and developer teams. This file describes public routes, approved positioning, benchmark retrieval-quality facts, source evidence, and machine-readable claim limits. Website: https://cortrix.ai/ Primary language: English Document version: 2026-09-06 Generated from public website content and commit-pinned source evidence Current Core release tag: https://github.com/cortrix/cortrix/tree/v1.0.0-rc.2 Current Core release commit: https://github.com/cortrix/cortrix/commit/d298d6b7e64c67551f567e7cddd514bae1adb57b Benchmark measured Core: https://github.com/cortrix/cortrix/commit/79a4eb17c62521338d1ac47a9749e6230e87e69b Benchmark public runner: https://github.com/cortrix/cortrix-benchmarks/commit/9490520c24a96ed97b80073ed3ebab096b80550b Benchmark RC2 evidence carrier: https://github.com/cortrix/cortrix-benchmarks/commit/5381a6a50795d546cf5b6dfc52fc892293033bf7 Benchmark immutable bundle: https://github.com/cortrix/cortrix-benchmarks/tree/4b94390c1d5f7be95065e7483362ec7f93774ed7/results/published/beir-four-corpus-cpu-2026-08-v1 Current maintainer reproduction guide: https://github.com/cortrix/cortrix-benchmarks/blob/5381a6a50795d546cf5b6dfc52fc892293033bf7/results/published/beir-four-corpus-cpu-2026-08-v1/reproduction.md Source/evaluation boundary: use the RC2 tag and resolved commit for current Core source documentation, the Docker Quick Start, the deeper maintainer evidence workflow, and adoption guidance. Benchmark scores remain tied to measured Core 79a4eb17..., public runner 9490520c..., and immutable bundle 4b94390c...; later release and license commits do not relabel the measured source. Canonical evidence states: Verified, Present in source - review required, Not established, Blocked, Future direction ## Public Routes - Home: https://cortrix.ai/ - Semantic storage guide: https://cortrix.ai/semantic-storage/ - For builders: https://cortrix.ai/for-builders/ - Benchmark: https://cortrix.ai/benchmark/ - Project updates: https://cortrix.ai/community/ - Human-readable llms.txt guide: https://cortrix.ai/docs/llms/ - Machine-readable scope: https://cortrix.ai/llms.txt - Agent-assisted local setup contract: https://cortrix.ai/agent-quickstart.txt - Current Core source: https://github.com/cortrix/cortrix/tree/v1.0.0-rc.2 - Docker Quick Start: https://github.com/cortrix/cortrix/blob/v1.0.0-rc.2/docs/QUICKSTART.md - Maintainer source-build evidence workflow: https://github.com/cortrix/cortrix/tree/v1.0.0-rc.2/examples/first-value-supportops - Stack-fit and adoption boundaries: https://github.com/cortrix/cortrix/blob/v1.0.0-rc.2/docs/adoption/stack-fit.md ## Agent workflow context - Cortrix explores semantic storage through retrieval, reviewable records, memory APIs, and source context. - Agent storage: data read and written by an agent workflow. - Context storage: source material and records retrieved to assemble model context. Storage alone does not decide what belongs in a prompt. - Memory storage: information retained for future interactions. Persistence and similarity do not establish validity, authorization, or safe sharing. - These explanations describe workflow needs, not additional shipped product capabilities. Existing source and security boundaries still apply. - Workflow overview: https://cortrix.ai/#semantic-storage - Definition and related concepts: https://cortrix.ai/semantic-storage/ - The conceptual guide is distinct from product documentation. Scott’s paper explainers are personal interpretations, not descriptions of papers implemented in Cortrix. ## Approved Public Positioning - Cortrix is an active pre-release, local-first, open-source semantic storage project for retrieval, reviewable records, memory APIs, and source context. - Cortrix is intended for AI-native engineers, agent builders, individual developers, independent teams, and developer teams. - Cortrix is designed to reduce repeated integration work; this website does not publish a measured maintenance-cost reduction. - Cortrix works with existing databases, tools, and agent workflows instead of attacking or replacing every database. - Benchmark content presents 16 maintainer-reported CPU retrieval cells across SciFact, NFCorpus, FiQA, and Quora, with each value attached to its arm and selected-query scope. - OpenAPI boundary: the current Core source commit contains the declared `api/openapi.yaml` file. File presence and identity do not establish runtime behavior, production readiness, or security readiness. - Present in source - review required: local server, MCP, Python SDK, built-in Agent, CUDA, OCR/parser paths, benchmark integration, and pgCortrix. - Not established: production readiness and security readiness. - Blocked: authentication, tenant/member/ACL/quota, RBAC and tenant-denial matrices, MEM02 extraction, and default log redaction. - Namespace, session, and user fields are logical scopes. They do not establish authentication, authorization, ACL, or tenant isolation. ## Benchmark Boundary - The public result bundle contains 16 maintainer-reported CPU retrieval results across SciFact, NFCorpus, FiQA, and Quora. The benchmark page displays four results: the best measured result for each corpus. Every value is an absolute nDCG@10 or Recall@10 result attached to a named arm and selected-query scope; the benchmark page does not present an RC1/RC2 comparison table. - SciFact, NFCorpus, and FiQA use their complete public corpora and every judged test query: 300, 323, and 648 respectively. Quora uses its complete 522,931-document corpus and the first 2,000 of 10,000 judged queries; it is not a full-query-set result. - All 16 cells completed every selected query with zero request failures and zero missing query text. - SciFact nDCG@10 / Recall@10: Dense + BM25 0.5942 / 0.7728; Embedding + Reranking 0.6184 / 0.8111; full-profile reranking 0.6206 / 0.8044; ingest-time LLM enrichment + reranking 0.6449 / 0.8214; Full Stack 0.7784 / 0.8752. - NFCorpus nDCG@10 / Recall@10: Dense + BM25 0.2991 / 0.1495; Embedding + Reranking 0.3121 / 0.1569; full-profile reranking 0.3091 / 0.1539; ingest-time LLM enrichment + reranking 0.3124 / 0.1589; Full Stack 0.3772 / 0.1768. - FiQA nDCG@10 / Recall@10: Dense + BM25 0.2540 / 0.3703; Embedding + Reranking 0.3125 / 0.4523; full-profile reranking 0.3156 / 0.4554. - Quora nDCG@10 / Recall@10: Dense + BM25 0.5003 / 0.7142; full-profile reranking 0.2031 / 0.3078; Full Stack 0.3039 / 0.3092. - The bundle measured only top_k=10. Deeper cutoffs are absent rather than estimated. - SciFact, NFCorpus, and FiQA arms were independently ingested. The bundle records an approximately 0.001-0.003 nDCG variation floor across separate ingests and an unexplained 0.0011 NFCorpus / 0.0027 FiQA residual. Quora arms share namespaces and are the only strictly controlled arm comparison. - FiQA carries a known score-fusion loss in the measured build. Quora's absolute values are below published dense-retrieval baselines. Read the bundle limitations before quoting any absolute value. - Timing remains in the scorecards for inspection but is not Website evidence of server performance, concurrency, throughput, or capacity. - Measured Core: https://github.com/cortrix/cortrix/commit/79a4eb17c62521338d1ac47a9749e6230e87e69b - Public runner: https://github.com/cortrix/cortrix-benchmarks/commit/9490520c24a96ed97b80073ed3ebab096b80550b - RC2 evidence carrier: https://github.com/cortrix/cortrix-benchmarks/commit/5381a6a50795d546cf5b6dfc52fc892293033bf7 - Immutable result bundle: https://github.com/cortrix/cortrix-benchmarks/tree/4b94390c1d5f7be95065e7483362ec7f93774ed7/results/published/beir-four-corpus-cpu-2026-08-v1 - Public method sources: https://github.com/cortrix/cortrix-benchmarks/blob/5381a6a50795d546cf5b6dfc52fc892293033bf7/benchmarks/beir-retrieval-quality/README.md ; https://github.com/cortrix/cortrix-benchmarks/blob/9490520c24a96ed97b80073ed3ebab096b80550b/benchmarks/beir-retrieval-quality/runner/run_benchmark.py ; https://github.com/cortrix/cortrix-benchmarks/blob/5381a6a50795d546cf5b6dfc52fc892293033bf7/benchmarks/beir-retrieval-quality/docs/methodology.md ; https://github.com/cortrix/cortrix-benchmarks/blob/5381a6a50795d546cf5b6dfc52fc892293033bf7/results/published/beir-four-corpus-cpu-2026-08-v1/reproduction.md - The bundle contains a manifest, summaries, 16 scorecards, reproduction instructions, and a complete SHA-256 inventory. These checks establish file identity and contract consistency, not an independent rerun or independent per-query recomputation. - Scope: retrieval accuracy and inspectable artifact identity only. Do not extrapolate these numbers to answer quality, memory quality, server performance, security, compliance, competitive ranking, agent workflow quality, or business outcomes. The RC2 release article separately reports the July-to-August cross-encoder comparison from the public Core and Benchmark releases. It compares July no_llm with August cross_encoder_rerank on the three common datasets, not Full Stack configurations; builds, configuration, and execution environments differ. The final release commit is distinct from the measured Core snapshot. ## Benchmark Evidence Format - Each cell records the corpus, arm, Recall@10, nDCG@10, selected-query scope, timing, run date, hardware, Cortrix source, runner identity, model content hashes, dataset archive hash, methodology, checksum inventory, and limitations. - The bundle is measured at top_k=10 only. Capped or deeper-cutoff runs require separate manifests and cannot be substituted for these cells. ## Claim Boundaries - Do not claim unannounced hosted or commercial offerings from this website. - Do not claim unannounced resilience, certification, data-sync, deduplication, customer-proof commitments, automatic self-learning retrieval, or specific deployment topology guarantees. - Do not treat public documentation or source presence as proof that a capability is Verified, shipped, production-ready, or security-ready. - Do not write attack content against other projects or tools. - Do not claim search placement, advertising conversion, AI overview inclusion, GEO success, or LLM recommendation outcomes. ## Social Links Visible - GitHub repository: https://github.com/cortrix/cortrix - LinkedIn Page: https://www.linkedin.com/company/cortrix/ - X: https://x.com/Cortrix_AI ## Project Participation - Core issues: https://github.com/cortrix/cortrix/issues - Core contribution guide: https://github.com/cortrix/cortrix/blob/v1.0.0-rc.2/CONTRIBUTING.md - Core security policy: https://github.com/cortrix/cortrix/blob/v1.0.0-rc.2/SECURITY.md - Core Code of Conduct: https://github.com/cortrix/cortrix/blob/v1.0.0-rc.2/CODE_OF_CONDUCT.md - Core maintainers and acknowledgment targets: https://github.com/cortrix/cortrix/blob/v1.0.0-rc.2/MAINTAINERS.md - Core DCO policy: https://github.com/cortrix/cortrix/blob/v1.0.0-rc.2/DCO - Published benchmark challenges: https://github.com/cortrix/cortrix-benchmarks/blob/main/CONTRIBUTING.md#challenge-a-published-result - Cortrix does not require a CLA. New commits require a DCO 1.1 Signed-off-by line; follow the applicable contribution guide before opening a pull request. ## License Boundary - Cortrix-authored material in v1.0.0-rc.2 is licensed under Apache-2.0. Historical v1.0.0-rc.1 release objects retain AGPL-3.0-only; third-party material retains its identified license. - Current Cortrix Benchmarks code is licensed under Apache-2.0. Documentation, CFF, schemas, curated results, and provenance remain under CC-BY-4.0; historical code revisions retain their original AGPL-3.0-only terms. - The cortrix.ai repository and Website content are publicly inspectable but do not grant a reuse license. Core and Benchmark licenses do not change the Website repository's copyright terms. ## Language Boundary - English is the canonical public language. - English remains the canonical source for machine-readable public boundaries. - English is the only public website language. Do not infer localized pages from this file. ## Blog - Blog index: https://cortrix.ai/blog/ - RSS feed: https://cortrix.ai/blog/rss.xml - Release update: Cortrix RC2: 24.4% higher FiQA recall, with Apache 2.0 — https://cortrix.ai/blog/cortrix-rc2-release/ — source release: https://github.com/cortrix/cortrix/releases/tag/v1.0.0-rc.2 — benchmark evidence: https://github.com/cortrix/cortrix-benchmarks/releases/tag/cortrix-v1.0.0-rc.2-evidence-v1 - Inside Cortrix: Vector is not semantic — https://cortrix.ai/blog/vector-is-not-semantic/ — scenario source: Order API Field Migration scenario, https://github.com/cortrix/cortrix-demos/tree/c3835a80026d22a8ed707511e9cbca6874c95e71/scenarios/order-api-field-migration - Research Note 01: DBpedia explained: why linked data still matters — https://cortrix.ai/blog/dbpedia-linked-data-explained/ — source paper: DBpedia: A Nucleus for a Web of Open Data (2007), https://doi.org/10.1007/978-3-540-76298-0_52 - Research Note 02: TransE explained: relationships beyond vector similarity — https://cortrix.ai/blog/transe-knowledge-graph-embeddings/ — source paper: Translating Embeddings for Modeling Multi-relational Data (2013), https://proceedings.neurips.cc/paper_files/paper/2013/hash/1cecc7a77928ca8133fa24680a88d2f9-Abstract.html - Research Note 03: RAG explained: external memory brings data governance back — https://cortrix.ai/blog/rag-external-memory-data-governance/ — source paper: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020), https://papers.nips.cc/paper_files/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html - Research Note 04: Reflexion explained: agent memory without weight updates — https://cortrix.ai/blog/reflexion-agent-memory-without-weight-updates/ — source paper: Reflexion: Language Agents with Verbal Reinforcement Learning (2023), https://proceedings.neurips.cc/paper_files/paper/2023/hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html - Research Note 05: HippoRAG explained: why RAG comes back to knowledge graphs — https://cortrix.ai/blog/hipporag-knowledge-graphs-for-rag/ — source paper: HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models (2024), https://proceedings.nips.cc/paper_files/paper/2024/hash/6ddc001d07ca4f319af96a3024f6dbd1-Abstract-Conference.html The Blog contains Cortrix product updates and Scott’s personal English interpretations of foundational papers. Research notes are not implementation notes for Cortrix. Release updates cite their source release and supporting evidence; research notes keep separate, visible paper citations. This metadata does not claim search placement, AI recommendation, or guaranteed discoverability.