Your AI finally remembers.

You explain your codebase to Claude. Architecture, constraints, patterns. Then you hit the context limit. New session. Claude forgets everything.

Session 1                         Session 2
───────────────────────────────────────────────────
You: "Our auth uses JWT..."       You: /catchup
Claude: *learns*                  Claude: "I remember:
         ↓                         - Auth uses JWT with refresh
     /wrap                         - Convention: validate at boundaries
         ↓                         What should we work on?"
    [Memory]  ────────────────►

Persistent, project-scoped memory with semantic search. Conventions, decisions, snippets — stored locally, never forgotten.

Option A: Automated with hive-mcp-cli (Recommended)

# Requires Go 1.21+
go install github.com/hive-agi/hive-mcp-cli/cmd/hive@latest
go install github.com/hive-agi/hive-mcp-cli/cmd/hive-setup-mcp@latest

# Register and let Claude guide setup
claude mcp add hive-setup --scope user -- hive-setup-mcp
claude
# Ask: "Help me setup hive-mcp"

Option B: Manual

export HIVE_MCP_DIR="$HOME/hive-mcp"
export BB_MCP_DIR="$HOME/bb-mcp"

git clone https://github.com/hive-agi/hive-mcp.git "$HIVE_MCP_DIR"
git clone https://github.com/hive-agi/bb-mcp.git "$BB_MCP_DIR"

claude mcp add hive --scope user -- "$HIVE_MCP_DIR/start-bb-mcp.sh"

claude mcp list | grep -q "hive" && echo "OK" || echo "FAILED"

ollama pull nomic-embed-text      # Local embeddings
docker compose up -d              # Chroma vector DB

Optional:

• Emacs 28.1+ for swarm vterm UI and buffer integration. See Emacs Configuration.
• Headless mode works without Emacs (but WIP for stability of headless - recommended to use Emacs as a dependency).

Each tool is a consolidated namespace with subcommands (e.g., memory add, memory query, kg traverse).

Claude ──MCP──► hive-mcp (Clojure)
                    │
                    ├── Memory (Chroma vectors + scoped entries)
                    ├── Knowledge Graph (DataScript, Datalevin, or Datahike)
                    ├── Swarm (lings + drones + hivemind)
                    ├── Workflows for AIs (building blocks for repeatable behavior)
                    └── Extension Registry (requiring-resolve stubs)
                            │
                            └──► [optional] your extensions (e.g., plug openclaw, langchain, whatever etc)

hive-mcp uses a plugin architecture based on the IAddon protocol with automatic classpath discovery. Creating a new addon takes one command:

clojure -Sdeps '{:deps {io.github.hive-agi/hive-mcp {:git/tag "vX.Y.Z" :git/sha "..."}}}' \
  -Tnew create :template hive-agi/addon :name com.example/my-addon

This generates a complete project with:

• IAddon protocol implementation (defrecord with 8 lifecycle methods)
• META-INF classpath manifest (auto-discovered at startup, zero core changes)
• Unit tests (12 tests covering lifecycle, tools, health)
• REPL-ready development setup

Addons are discovered via META-INF/hive-addons/*.edn manifest files on the classpath (same pattern as Java's ServiceLoader). Manifests declare dependencies, and addons are loaded in topological order. No changes to hive-mcp core code needed.

See the Creating Addons wiki guide and ADR-0007 for architecture details.

Cartography snippets (code-chunked embeddings for carto_callers / carto_definition / carto_callees) live in a dedicated vector store — separate from the general :memory-store — so recall tuning stays independent.

Default config (already wired in docker-compose.yml as service qdrant-carto):

{:services
 {:qdrant-carto {:mode :local
                 :host "localhost"
                 :port 6333
                 :collection "carto-snippets"
                 :embedding {:provider :ollama
                             :model "nomic-embed-code"}}
  :carto-store  {:backend :qdrant-carto}}}

User override example (~/.config/hive-mcp/config.edn):

{:services
 {:qdrant-carto {:host "qdrant.internal"
                 :port 6333
                 :collection "my-carto"
                 :embedding {:provider :ollama
                             :model "nomic-embed-code"}}
  :carto-store  {:backend :qdrant-carto}}}

If :carto-store is omitted, hive-mcp.config.resolve/resolve-carto-store logs a warning and falls back to :memory-store — the system keeps booting, but cartography writes land in the shared store. Production deployments should declare :carto-store explicitly.

Bring up the dedicated instance:

docker compose up -d qdrant-carto

See CLAUDE.md for project conventions, tool patterns, and memory usage guidelines.

AGPL-3.0-or-later