neurodock-mcp-cognitive-graph 0.0.3

NeuroDock cognitive graph MCP server — persistent entity memory and recall.

neurodock-mcp-cognitive-graph

Persistent entity memory for the NeuroDock substrate, exposed as an MCP server. Externalises the user's working memory of people, projects, decisions, and concepts so a hyperfocused or context-switched session does not start from zero.

This is v0.0.1 — the first working slice. Vector recall, fastembed embeddings, and sqlite-vec are deferred to v0.0.2; see CHANGELOG.md.

Quickstart

# From the workspace root.
uv sync --all-packages

# Start the server on stdio (Claude Desktop / Claude Code MCP config).
uv run neurodock-mcp-cognitive-graph

The server stores its graph at ~/.neurodock/cognitive-graph.sqlite by default. Override with the NEURODOCK_GRAPH_DB_PATH environment variable.

Tools (v0.0.1)

Tool	Purpose
recall_entity(name_or_alias)	Look up a person/project/decision/concept. Returns the entity, its facts (capped at 500), first-degree neighbours (capped at 20), and a resolution diagnostic.
record_fact(subject, predicate, object, source?, confidence?)	Persist a typed-edge fact. Auto-creates entities by (type, name). Enforces the v0.1.0 eight-predicate vocabulary.
recall_decisions(project, since?)	Return decisions for a project, ordered by date desc, capped at 200.
weekly_rollup(project?)	Server-templated activity summary for the trailing seven UTC days. No LLM call (vendor boundary).

The full input/output contract is in schemas/. The Pydantic v2 models in src/neurodock_mcp_cognitive_graph/types.py mirror those schemas.

Predicate vocabulary

The eight predicates in v0.1.0 are: mentioned_in, decided_in, reports_to, depends_on, resolved_by, blocked_by, tagged, belongs_to. Unknown predicates raise PREDICATE_NOT_IN_VOCABULARY. Extension predicates land via the v0.2 type_extensions mechanism — see ADR 0002 — cognitive-graph tool design.

Privacy

• Local-first. No network access. No telemetry.
• source strings are stored verbatim and never fetched by the server.
• User content (entity names, decision titles, fact text) is never logged. Error logs include only the structured error code.

Storage

SQLite. One file. Migrations live as numbered .sql files in src/neurodock_mcp_cognitive_graph/migrations/ and are applied on first connect. The schema is described in migrations/0001_init.sql.

Architecture

src/neurodock_mcp_cognitive_graph/
├── server.py            FastMCP wiring + CLI entrypoint
├── types.py             Pydantic v2 models (mirror of the JSON Schemas)
├── config.py            Resolves the SQLite path
├── clock.py             SystemClock / FixedClock
├── errors.py            ToolError envelope
├── resolution.py        Entity name/alias cascade (exact, alias only in v0.0.1)
├── rollup.py            Heuristic decision/blocker/next-action assembly
├── storage/
│   ├── base.py          Storage Protocol + row dataclasses
│   ├── memory.py        InMemoryStorage (used by tests)
│   └── sqlite.py        SQLiteStorage (production backing)
├── tools/
│   ├── recall_entity.py
│   ├── record_fact.py
│   ├── recall_decisions.py
│   └── weekly_rollup.py
└── migrations/
    └── 0001_init.sql

ADR pointers

• ADR 0002 — cognitive-graph tool design — the design rationale, vocabularies, output shapes, and the three open questions whose v0.0.1 resolutions are documented in CHANGELOG.md.
• ADR 0001 — chronometric tool design — the cross-cutting precedents (ISO 8601 with offsets, enums for coarse signals, structured errors) that this package inherits.

Tests

uv run pytest packages/mcp-cognitive-graph/tests/ -v

25 tests covering: per-tool unit tests, a FastMCP protocol-conformance suite that exercises every tool and validates the response against the JSON Schema using jsonschema, and an end-to-end test that runs against a real file-backed SQLite store.