v3.0.0

• feat: redesign unified graph visualization UX
• docs: update hero image alt text for v2.6.0 screenshots
• docs: update hero images with v2.6.0 unified graph screenshots

Full Changelog: v2.6.0...v3.0.0

v2.6.0

• fix: add HF model cache to release workflow
• fix: add sentence-transformers and networkx to dev deps for CI
• fix: move model pre-download after pip install, allow soft failure
• fix: cache HuggingFace model in CI to prevent download failures
• fix: ruff format http_standalone.py
• release: v2.6.0 — remove memory dashboard, global memory dedup fix, cleanup

Full Changelog: v2.5.5...v2.6.0

v2.5.2

• release: v2.5.2 — schema resilience, lint fixes, all backends aligned
• fix: ruff format + remove unused variable to pass CI lint
• fix: consistent per-statement schema init across PG, SQLite, and Docker
• feat: add cortex-remember-global and cortex-recall-global skills
• feat: global memories — cross-project knowledge with auto-detection and visualization
• fix: split DDL into individual statements to prevent cascading schema failures

Full Changelog: v2.5.1...v2.5.2

v2.5.1

• release: v2.5.1 — version bump (2.5.0 tarball burned on PyPI)

Full Changelog: v2.5.0...v2.5.1

v2.5.0

• fix: align package.json and safeskill-report.json to v2.5.0
• release: v2.5.0
• fix: 5 user-reported bugs — path validation, hierarchical recall, causal chain, narrative, seed
• fix: persistent memory across Docker container restarts
• fix: use shields.io badge for SafeSkill — their badge API is broken
• fix: SafeSkill badge uses dynamic API from published npm package
• fix: Docker entrypoint registers MCP in .claude.json, strips host config
• fix: badge links to local SafeSkill report
• docs: add SafeSkill 94/100 scan report, remove stale badge link
• fix: add minimal package.json for SafeSkill scanner, update badge to 94/100
• fix: remove legacy Node.js files flagged by SafeSkill security scan
• Merge pull request #2 from OyaAIProd/safeskill-scan-1774780323456
• Add SafeSkill security badge (80/100)
• fix: Docker runs Claude Code inside container with stdio MCP (no HTTP bridge)

What's Changed

• Add SafeSkill security badge (80/100 — Passes with Notes) by @OyaAIProd in #2

New Contributors

• @OyaAIProd made their first contribution in #2

Full Changelog: v2.4.1...v2.5.0

v2.4.1 — Fix UI path resolution for plugin installs

• fix: release workflow installs postgresql extras for test isolation
• release: v2.4.1 — fix UI path resolution for uv/plugin installs
• fix: UI files not found when installed via uv plugin cache
• docs: add Docker install path to README, update plugin descriptions
• feat: Docker image with PostgreSQL, pgvector, and Claude Code CLI
• fix: bundle UI in wheel, standalone HTTP servers survive MCP shutdown (v2.4.0)

Full Changelog: v2.3.0...v2.4.1

v2.4.0

What's Changed

Bug Fixes

• UI files bundled in package — memory-dashboard.html, unified-viz.html, and methodology-viz.html were missing from pip/uv installs. Added hatch force-include to bundle ui/ into the wheel, plus get_ui_root() to resolve both package and dev layouts.

• Visualization servers survive session end — HTTP servers previously ran as daemon threads inside the MCP process and died when Claude's session ended, leaving a dead browser page. Now spawn as detached subprocesses (start_new_session=True) with their own DB connections and 10-minute idle self-termination. No human intervention needed — call the tool, browser opens, server stays alive.

• Tree-sitter fallback hardened — ast_parser now gracefully falls back to regex parsing when tree-sitter-language-pack is not installed, instead of crashing with ModuleNotFoundError.

Testing

• Removed all test skips — tests adapt to available dependencies (tree-sitter, networkx, PostgreSQL) instead of skipping
• Fixed test isolation — stats and checkpoint tests no longer assume empty DB, spell benchmark tests fall back to direct store queries when semantic recall is unavailable
• All CI jobs passing on Python 3.10–3.13

v2.3.0 — Codebase Intelligence + SQLite Fallback

What's New

Codebase Intelligence (replaces GitNexus)

Index your codebase directly into Cortex memory — no external tools needed.

codebase_analyze(directory="/path/to/project")

• Tree-sitter AST parsing for Python, TypeScript, Go, Swift, Rust (regex fallback)
• Cross-file import resolution — resolves module names to actual files
• Type-reference resolution — for Swift/Go implicit cross-file type usage
• Class-method binding — methods scoped to parent class
• Inheritance tracking — extends edges in the knowledge graph
• Community detection — Louvain clustering via networkx
• Impact analysis — upstream/downstream BFS
• Incremental — SHA-256 hash, only re-processes changed files
• Works on both PostgreSQL and SQLite (Cowork compatible)

Pipeline → Memory Integration

Pipeline events flow back to Cortex memory. Each stage stores observations with structured tags for cross-session recall.

SQLite Fallback

Cortex runs in sandboxed environments without PostgreSQL. SQLite + sqlite-vec provides the same API with automatic fallback.

Spell Alteration Benchmark

Needle-in-a-haystack: ingest 1.5M tokens as 3000+ memories, replace 2 terms, test detection.

• Test A: Spot fakes in 3372 memories — PASS
• Test B: Compare original vs altered, identify correct pairing — PASS

Stats

• 35 MCP tools (+1 codebase_analyze)
• 2000+ tests passing across Python 3.10-3.13
• 7 benchmarks

Install

```bash
claude plugin marketplace add cdeust/Cortex
claude plugin install cortex
```

Optional:
```bash
pip install neuro-cortex-memory[codebase] # tree-sitter + networkx
pip install neuro-cortex-memory[postgresql] # psycopg + pgvector
```

v2.1.0

Full Changelog: https://github.com/cdeust/Cortex/commits/v2.1.0