DataMind is a unified retrieval agent that wires six knowledge capabilities into one Claude-powered tool registry:

Capability	What it does	Default backend
KB (RAG)	Semantic + lexical document retrieval	Chroma + BM25 (Reciprocal Rank Fusion)
Graph	Entity lookup and multi-hop traversal	NetworkX (JSON-persisted)
Database	Natural language → SQL query	SQLAlchemy (SQLite / MySQL / Postgres)
Skills	Markdown SOPs + safe code skills (calculator, unit conversion, ...)	.claude/skills/<name>/SKILL.md
Memory (v0.3 scope-typed)	Short-term buffer + SQLite long-term with cosine recall; three scopes (global / profile / session) for multi-tenant isolation	SQLite + embeddings
Hooks (v0.3 sandboxed)	Intercept every tool call: Allow / Deny / AskUser / Rewrite; built-in destructive-SQL gate, path allow-list, tamper-evident audit log	HookChain + Merkle audit

The agent loop picks tools on its own, recovers from errors, and streams output throughout. You don't hard-code which question routes to which capability.

What changed since v0.1

v0.1 was a LlamaIndex FunctionAgent with everything wired into a global AppState. v0.3 keeps the same six capabilities but reshapes the system around three structural ideas:

• Protocol + Registry kernel. Each capability is defined by a small Protocol; concrete implementations register themselves under a name. Adding a new SQL dialect / embedding provider / retriever is a single-file change with zero impact on the core.
• Pluggable agent loop. Two interchangeable backends (native over the Anthropic SDK, sdk over claude-agent-sdk + CCR) share one tool registry and one SSE event format. Switch with one env variable.
• Real SSE streaming, not the v0.1 "character-sliced" simulation. Per-request RequestContext propagates trace_id through every tool call and audit record.

v0.3 adds two flagship features on top: scope-typed Memory (multi- tenant isolation without a separate deployment per customer) and hook-interposed dispatch (destructive-SQL confirmation, path allow-list, tamper-evident audit log).

The legacy v0.1 entry points (main.py / server.py / modules/) remain in the repo for side-by-side comparison and are still runnable.

Repository layout

datamind/
├── agent/              # agent loop + system prompt + capability assembly
├── capabilities/
│   ├── embedding/      # OpenAI-compatible + HuggingFace
│   ├── kb/             # Chroma + simple/multi_query/hybrid retrievers
│   ├── graph/          # NetworkX graph store
│   ├── db/             # SQLAlchemy + SQLite/MySQL dialects + NL2SQL
│   ├── memory/         # short-term rolling + SQLite long-term (scope-typed)
│   ├── skills/         # SKILL.md loader + code skills
│   ├── ingest/         # conversational ingest into KB / DB / Graph
│   └── hooks/          # PathAllowlist / DestructiveSql / AuditLog
├── core/               # Protocol / Registry / Config / Logging / Tools / Hooks
├── scripts/            # one hello_*.py per capability — real smoke tests
├── cli.py              # `python -m datamind ...`
├── server.py           # FastAPI + real SSE
└── tests/              # 133 passing tests, no network

Every capability ships with a hello_<cap>.py that runs against a real gateway. Start with Install & run.