Local-first RAG for code & docs, with long-term memory for AI agents. BM25, vector, GraphRAG, and hybrid search over your codebase and documentation — plus session memory, a temporal knowledge graph, and git/LSP-aware indexing. Use it from the CLI, over MCP, or as a Claude Code plugin. Runs fully local on Ollama, or with cloud LLMs.

Pick the path that matches how you'll use BrainPalace. All paths share the same prerequisites and end with the same brainpalace CLI on your PATH.

Richest UX — 30 slash commands, 3 agents, 2 skills. The setup wizard inside Claude Code installs the CLI + server, configures provider keys, initialises the project, starts the server, and runs the first index — all from one slash command.

1. Install the plugin:

   claude plugins install github:bxw91/brainpalace

2. Run the setup wizard in Claude Code:

   /brainpalace-setup
   

Full Claude Code reference: docs/PLUGIN_GUIDE.md.

Use this if you want brainpalace as a command-line tool, or if you want to connect an MCP-capable editor (Cursor, VS Code Copilot, Cline, Continue, Kilo Code, Zed). One command does both — it will ask which MCP client to wire (or "none" for CLI-only) along the way. Nothing runs until you answer.

curl -sSL https://raw.githubusercontent.com/bxw91/brainpalace/stable/scripts/setup.sh | bash

That's it.

Got more projects to index later? The binary is installed once per machine; every new project needs three commands (brainpalace init --start --watch auto, then brainpalace index .). Full how-to: docs/INSTALL.md → Adding more projects.

CI / no-TTY, step-by-step manual install, low-level install.sh flags, install for other AI runtimes (Codex, OpenCode, Gemini CLI), Windows / WSL2 notes — all in docs/INSTALL.md.

BrainPalace indexes your codebase and documentation, then exposes the resulting search over multiple interfaces so any AI assistant — or you yourself — can answer questions against it. On top of plain retrieval it keeps long-term memory: curated facts, captured coding-session summaries and decisions, and a temporal knowledge graph that tracks how those decisions supersede each other over time. Local-first by default (Ollama), with optional cloud providers for embeddings and summarisation.

• Hybrid search — BM25 + vector + GraphRAG, fused per query (hybrid, multi) or selectable per call (bm25, vector, graph).
• Session intelligence — capture AI-coding sessions into curated memory (remember/recall, markdown source-of-truth), searchable summaries + decisions, and a typed knowledge graph (Decision / Error / File / …). Cross-session linking supersedes stale decisions and promotes durable ones into memory. See SESSION_INDEXING.
• Persistent graph backend — opt-in store_type: sqlite with temporal validity (per-edge validity windows, invalidate, timeline); scales past the in-memory default. See GRAPHRAG_GUIDE.
• Time-decay ranking — newer chunks rank higher (configurable half-life).
• Git-history indexing — commit messages + diff stats as a searchable source, bridging why ↔ what. See GIT_HISTORY.
• LSP cross-references (opt-in) — typed calls/extends/implements symbol graph from a real language server. See LSP_INTEGRATION.
• AST-aware code chunking — tree-sitter for Python, TypeScript, JavaScript, Java, Kotlin, C, C++, C#, Go, Rust, Swift.
• LLM code summaries — optional AI-generated per-chunk descriptions to lift semantic recall on code.
• GraphRAG — entity + relationship extraction. Dependency-aware queries: "what calls X", "modules importing Y", "classes extending Z".
• Cross-encoder reranking — opt-in two-stage retrieval for higher precision on the top-k.
• .gitignore-aware indexing + watching — every project .gitignore (nested files, negation patterns) honoured at index and watch time.
• File watcher — per-folder, debounced, post-enqueue cooldown. Default OFF, opt-in per folder (auto).
• Multi-instance — one server per project, automatic port allocation, .brainpalace/runtime.json discovery. Helpers: whoami, status --all, stop --url.
• URL auto-discovery — CLI walks up from CWD to the owning server. Works correctly in mono-repos.
• Incremental indexing — manifest + SHA-256; only changed files re-embed; chunk eviction tracks deletes.
• Embedding cache — TTL 3600 s, hit-rate tracked. Cuts provider cost on reindex.
• Pluggable providers — embeddings (OpenAI · Cohere · Ollama), summarisation (Anthropic · OpenAI · Gemini · Grok · Ollama). Fully local mode via Ollama for both.

Run completely offline with Ollama for both embeddings and summarisation:

/brainpalace-providers
# Pick Ollama for both

Or via the CLI: docs/PROVIDER_CONFIGURATION.md.

The plugin ships 30 slash commands, 3 agents, and 2 skills. Full reference: docs/PLUGIN_GUIDE.md.

brainpalace/
├── brainpalace-plugin/                     # Claude Code plugin
│   ├── commands/                            # 30 slash commands
│   ├── agents/                              # 3 intelligent agents
│   ├── skills/                              # 2 context skills
│   └── templates/                           # mcp-config-claude-code.json + sessionstart hook
├── brainpalace-server/                     # FastAPI backend (REST API)
├── brainpalace-cli/                        # CLI + Python SDK + MCP shim
│   └── brainpalace_cli/
│       ├── commands/                        # CLI subcommands incl. `mcp`
│       ├── mcp_server/                      # Opt-in MCP stdio shim
│       └── client/                          # Python SDK
└── docs/                                    # User + developer docs

• Install (alternative paths) — manual / CI / other AI runtimes / low-level flags
• Quick Start — first-run walkthrough
• MCP Setup — per-client config for non-Claude-Code AI clients
• Plugin Guide — full Claude Code plugin reference
• User Guide — CLI usage and feature reference

• API Reference — REST API documentation
• Configuration — config.yaml options
• Provider Configuration — embedding + summarisation provider setup
• Changelog — per-version notes

• Architecture Overview — components, data flow
• GraphRAG Guide — knowledge-graph features
• Code Indexing — AST-aware chunking
• Deployment — local + production deployment
• Developer Guide — monorepo layout, sub-modules, contributing

git clone https://github.com/bxw91/brainpalace.git
cd brainpalace
task install
task before-push      # full quality gate — mandatory before merge

Full setup and contribution workflow: docs/DEVELOPERS_GUIDE.md.

• Server: FastAPI + Uvicorn
• Vector Store: ChromaDB (HNSW, cosine similarity)
• BM25 Index: LlamaIndex BM25Retriever
• Graph Store: LlamaIndex SimplePropertyGraphStore (JSON) or SQLite (persistent, temporal)
• Embeddings: OpenAI · Cohere · Ollama
• Summarisation: Anthropic · OpenAI · Gemini · Grok · Ollama
• AST Parsing: tree-sitter (10+ languages)
• CLI: Click + Rich
• MCP: Anthropic mcp SDK (stdio transport)
• Build: Poetry

PRs land on stable. Before pushing, task before-push must pass. See docs/DEVELOPERS_GUIDE.md for monorepo layout, test commands, and release discipline.

MIT — see LICENSE and NOTICE for details.

• Releases
• Issues