Give your agents
a brain.
Three open-source products for Claude Code that actually ship: Cortex persistent memory, an autonomous pipeline that takes findings to pull requests, and a PRD generator that verifies every claim. Built on neuroscience research, not guesswork.
Cortex — Persistent memory for Claude Code
97.8% recall on LongMemEval. 41 scientific citations. 18 zetetic agents with transactive memory. No GPU required.
/cortex-visualize.You don't manage memory — Cortex does. Session hooks capture decisions, protect important context, and prime related memories via spreading activation. Between sessions, a dream cycle decays noise and consolidates knowledge. Every agent starts briefed with prior work and team decisions via transactive memory (Wegner 1987). The neural graph visualization is built in.
Three products. One platform.
Each one solves a real problem. Each one is open source. Each one ships.
Cortex
Persistent memory that learns
97.8% recall on LongMemEval. 18 zetetic agents with transactive memory. PostgreSQL + pgvector. Built on 41 neuroscience papers. Includes neural graph visualization and memory monitoring out of the box.
- Session hooks capture & protect decisions
- Dream cycle consolidates between sessions
- Spreading activation primes related context
- 33 MCP tools, 14 skills, 7 hooks
- No GPU required
Pipeline
Findings to verified PRs, autonomously
11-stage autonomous pipeline with 66 MCP tools, 64 deterministic HOR rules, and 5 science-backed git analytics. Zero LLM-judges-LLM — every verification is algorithmic. Integrated with Cortex memory and the PRD Generator out of the box.
- 11 stages: health check to verified PR
- 3 MCP servers, 66 tools (49 + 17)
- 64 HOR rules across 10 categories
- 5 verification + 5 prompting algorithms
- Cortex memory compounds across runs
- Tree-sitter for 13 languages
PRD Generator
Verified specs, not just generated
15 thinking strategies, 6 verification algorithms, 8 PRD types. Generates verified Product Requirements Documents with claim-by-claim verification, JIRA tickets, test cases, and audit trails. Most PRD tools generate. This one verifies.
- 15 thinking strategies
- 6 independent verification algorithms
- 9-file PRD output (overview to verification)
- Multi-judge consensus scoring
- JIRA-ready ticket generation
Need a turnkey solution?
Solution Packs ship domain-specific MCP servers, pre-configured agents, and Cortex memory seeds for 6 industries. Plug into Claude Code — agents have expertise on day one.
Compliance & Governance
SOC 2, GDPR, HIPAA evidence collection & audit prep
Engineering Intelligence
Codebase analysis, PRD pipeline, dependency-aware decomposition
Legal & Contract
Privilege-aware e-discovery, contract review, Bates numbering
Financial Intelligence
Stripe analytics, SaaS expense optimization, anomaly investigation
Healthcare Operations
HIPAA-compliant intake, eligibility, clinical coding
Customer Intelligence
Sentiment analysis, content strategy, competitive intelligence
Why these three?
Built on research. Tested in production. Open sourced because they actually work.
Memory compounds
Tools without memory restart from zero every session. Cortex remembers what worked, what failed, and why — across sessions, projects, and team members. Run 1: agent explores. Run 100: agent anticipates.
Verification beats generation
Most AI tools generate output and call it done. The PRD Generator verifies every claim with 6 independent algorithms and multi-judge consensus. The Pipeline runs quality gates at every stage. Output you can ship.
Autonomy with accountability
The Pipeline takes a finding to a pull request without human intervention. But every decision is traceable: PRD claims, code changes, verification results, benchmarks. You can audit what the agent did and why.
Get started in 2 minutes
All three products install via Claude Code marketplace.
Cortex
claude plugin marketplace add cdeust/Cortex
claude plugin install cortex
Then run /cortex-setup-project. Handles PostgreSQL, embeddings, hooks — everything.
Pipeline
claude plugin marketplace add cdeust/ai-architect-mcp
claude plugin install ai-architect
Then run /ai-architect:run-pipeline with your finding. 11 stages, autonomous, Cortex-integrated.
PRD Generator
claude plugin marketplace add cdeust/ai-prd-generator-plugin
claude plugin install ai-prd-generator
Then run /ai-prd-generator. Auto-detects PRD type and verifies every claim.
All three are free and open source. Requires Claude Code or Cowork subscription from Anthropic.
FAQ
Cortex is persistent memory for Claude Code. Without it, every Claude Code session starts from zero — no memory of past decisions, failed approaches, or team context. With Cortex, your agents learn and compound knowledge across sessions.
It scores 97.8% Recall@10 on LongMemEval (ICLR 2025) and 92.6% on LoCoMo, runs on PostgreSQL + pgvector with FlashRank cross-encoder reranking (no GPU), and includes 18 specialized zetetic agents with transactive memory. The neural graph visualization is built in — launch with /cortex-visualize.
The Pipeline orchestrates an 11-stage autonomous workflow from a finding (bug, feature request, research paper) to a verified pull request. Claude Code can do individual steps; the Pipeline coordinates them: health check, discovery with typed codebase graph, compound impact scoring, integration design, PRD generation, 10-dimension interview gate, 7 verification algorithms, one-worker-per-file implementation, 64 HOR rules, benchmarks, deployment, and PR + Cortex memory storage.
Three MCP servers (66 tools total), zero LLM-judges-LLM, 544 tests passing. Integrated with Cortex and the PRD Generator. Free and open source.
The PRD Generator verifies every claim with 6 independent algorithms and multi-judge consensus. It generates a 9-file PRD package (overview, requirements, acceptance criteria, technical design, user stories, test plan, roadmap, JIRA tickets, verification report) and iterates until verification passes.
Most PRD tools stop at generation. This one tells you whether the PRD is internally consistent, technically feasible against your actual codebase, and testable.
Solution Packs are turnkey domain solutions for Claude Code. Each pack includes: a domain-specific MCP server (the hard part), 2 pre-configured agents with workflows, a Cortex memory seed with industry knowledge, and CLAUDE.md domain rules.
Six packs available: Compliance & Governance, Engineering Intelligence, Legal & Contract, Financial Intelligence, Healthcare Operations, Customer Intelligence. Plug into Claude Code and your agents have domain expertise on day one.
Yes — they're designed as a stack. Cortex provides persistent memory. The Pipeline uses Cortex to remember past pipelines and decisions. The PRD Generator (used at Stage 5 of the Pipeline) stores verified PRDs in Cortex for future reference. Solution Packs ship Cortex memory seeds and integrate with both.
You can also use any product standalone. Cortex is the most popular standalone install.
- Cortex: Python 3.10+, PostgreSQL 15+ with pgvector. Any OS.
- Pipeline: Claude Code, Node.js 18+. Cross-platform.
- PRD Generator: Claude Code only. Cross-platform (macOS, Linux, Windows).
- Solution Packs: Python 3.10+, plus pack-specific dependencies (PyGithub, boto3, stripe, etc.).
All require an active Claude Code or Cowork subscription from Anthropic.
Cortex, Pipeline, and PRD Generator are free and open source. Install via Claude Code marketplace.
Solution Packs have a tiered model: Open Source (free), Pro ($99/mo, 1 pack), Team ($299/mo, 3 packs + shared memory), Managed ($2K+/mo, all packs + custom MCP development). See Solutions pricing.
Get in Touch
Bug report, feature request, or need help integrating?