Cortex

Cognitive knowledge system with formal ontology, reasoning, and intelligence serving.

Cortex captures knowledge objects (decisions, lessons, fixes, sessions, research, ideas), classifies them with an OWL-RL ontology, discovers relationships, reasons over the graph, and serves intelligence through hybrid retrieval.

Install

pip install abbacus-cortex[embeddings]
cortex setup

The setup wizard configures everything: LLM provider, embeddings, dashboard password, background services, and Claude Code registration.

Lightweight (no PyTorch — wizard offers to install embeddings):

pip install abbacus-cortex
cortex setup

Homebrew:

brew install abbacusgroup/tap/abbacus-cortex
cortex setup

From source:

git clone https://github.com/abbacusgroup/Cortex.git && cd Cortex
uv sync --extra embeddings
cortex setup

Docker:

docker compose up -d

For non-interactive installs (CI, scripts): cortex setup --auto uses environment variables.

Quick Start

cortex capture "Fix: Neo4j pool exhaustion" --type fix --content "Root cause was..."
cortex search "Neo4j"
cortex list
cortex context "Neo4j"
cortex dashboard                    # web UI at http://localhost:1315

Configuration

cortex setup writes configuration to ~/.cortex/.env. You can also edit it directly:

CORTEX_LLM_PROVIDER=anthropic
CORTEX_LLM_MODEL=claude-sonnet-4-20250514
CORTEX_LLM_API_KEY=sk-...
CORTEX_EMBEDDING_MODEL=all-mpnet-base-v2

See .env.example for all options.

CLI Commands

MCP Tools

22 tools for AI agent integration. Localhost-bound HTTP exposes all; non-localhost binds expose only the public set.

Public: cortex_search, cortex_context, cortex_dossier, cortex_read, cortex_capture, cortex_link, cortex_feedback, cortex_graph, cortex_list, cortex_classify, cortex_pipeline

Admin (localhost only): cortex_status, cortex_synthesize, cortex_delete, cortex_reason, cortex_query_trail, cortex_graph_data, cortex_list_entities, cortex_export, cortex_safety_check, cortex_debug_sessions, cortex_debug_memory

Architecture

Cortex runs as a single MCP HTTP server that owns the graph store. Claude Code, the dashboard, the CLI, and the REST API are all HTTP clients of that one server.

┌───────────────┐    ┌────────────┐    ┌─────────────┐
│ Claude Code   │    │  Dashboard │    │     CLI     │
│ (MCP client)  │    │ (browser)  │    │  (terminal) │
└───────┬───────┘    └─────┬──────┘    └──────┬──────┘
        │                  │                  │
        │ HTTP JSON-RPC    │ HTTP MCP         │ HTTP MCP (default)
        │                  │                  │ direct (--direct)
        ▼                  ▼                  ▼
        ┌──────────────────────────────────────┐
        │   cortex serve --transport mcp-http  │
        │   (canonical MCP HTTP server)        │
        │   PID-locked owner of graph.db       │
        └──────────────────────────────────────┘
                          │
                          ▼
            ┌─────────────────────────────┐
            │  ~/.cortex/                 │
            │    graph.db   (Oxigraph)    │
            │    cortex.db  (SQLite WAL)  │
            └─────────────────────────────┘

• Ontology: OWL-RL formal ontology with 8 knowledge types and 8 relationship types
• Storage: Oxigraph (RDF/SPARQL) + SQLite (FTS5/BM25) dual-write
• Pipeline: Classify → Extract entities → Link → Enrich → Reason
• Retrieval: Hybrid keyword + semantic + graph-boosted ranking
• Serving: 5 presentation modes (briefing, dossier, document, synthesis, alert)
• Transports: MCP (stdio + HTTP), REST API, Web Dashboard

Service Management

# Install both MCP server and dashboard as background services
cortex install

# Install only the MCP server
cortex install --service mcp

# Remove all services
cortex uninstall

On macOS, this creates LaunchAgent plists (auto-start on login, auto-restart on crash).
On Linux, this creates systemd user units.

Raw templates are available in deploy/ for manual setup.

--direct escape hatch

By default, CLI commands route through the running MCP server. If the server is down:

cortex --direct list               # bypass MCP, open store directly
cortex --direct pipeline --batch   # required for bulk SQL operations

Bootstrap commands (init, setup, import-v1, import-vault) always run directly.

Docker

docker compose up -d
# Server at http://localhost:1314

Troubleshooting

Crashed MCP server

If the MCP server is killed hard, stale lock files auto-recover on next start. For manual cleanup:

cortex doctor unlock              # normal cleanup
cortex doctor unlock --dry-run    # report only
cortex doctor unlock --force      # bypass live-holder check

Log management

cortex doctor logs                # show log file sizes and status
cortex doctor logs --tail 20      # last 20 lines
cortex doctor logs --rotate       # rotate log files (safe while running)

Claude Code session staleness

After restarting the MCP server, restart Claude Code to clear its stale session ID. claude --resume restores your conversation. The dashboard and CLI do not have this issue.

Knowledge Types

decision, lesson, fix, session, research, source, synthesis, idea

Relationship Types

causedBy, contradicts (symmetric), supports, supersedes (transitive), dependsOn, ledTo (inverse of causedBy), implements, mentions

Privacy

Cortex stores all data locally. No telemetry, no analytics, no phone-home. If you configure an LLM provider (via CORTEX_LLM_API_KEY), object content may be sent to that provider for classification and reasoning. Embeddings are computed locally by default using sentence-transformers.

License

Copyright (c) 2026 Abbacus Group. Licensed under the MIT License.

Trademark Notice

Cortex is a project of Abbacus Group and is not affiliated with any other product named Cortex.