Neotoma with Claude

Start with evaluation. Once Neotoma has been evaluated, installed if needed, and activated with your first stored data, reliable Claude usage requires four more steps: connect the MCP server, set tool permissions, create a Project, and add routing instructions. Each step builds on the previous one.

1. Connect the Neotoma MCP server
2. Set tool permissions to “Always allow”
3. Create a Claude Project for Neotoma-backed work
4. Add project instructions that route queries to Neotoma

Choose a connection method based on your setup. Claude Desktop uses local stdio; claude.ai uses remote MCP over HTTPS.

After connecting, verify the connector appears in Settings → Connectors and shows a “Connected” state. You should see Neotoma listed with its tools.

By default Claude may ask for permission before using connector tools, which breaks automatic retrieval. Set all Neotoma tools to always run without prompting:

1. Open Settings → Connectors
2. Click Neotoma in the connector list
3. Under Tool permissions, set the dropdown for Other tools to Always allow

This ensures Claude can call Neotoma tools (retrieve, store, search) without interrupting you for approval each time.

Claude Projects let you attach persistent instructions that apply to every chat inside the project. This is the only way to give Claude standing orders on claude.ai without repeating them each conversation.

1. In the Claude sidebar, click Projects → Create project
2. Name it something like Neotoma workspace (or whatever fits your use case)
3. Open the project. All new chats started inside it will inherit its instructions

Open your project's settings and paste the following into the Custom instructions field. This tells Claude to route all factual retrieval through Neotoma and refuse to answer from internal memory alone.

For any request that depends on stored facts, history, lists, counts, records,
timelines, prior work, or "what do you know / show me / list / summarize":

1. Call Neotoma MCP tools before answering.
2. Do not answer from Claude memory or chat history alone.
3. If Neotoma is unavailable or tool call fails, respond:
   "Neotoma retrieval failed: <error>. Please check the connector."
4. If no matching records are found, say so explicitly and stop.
5. Every factual claim must be traceable to Neotoma tool output from this turn.

This is a fail-closed policy: if Neotoma is unreachable, Claude says so instead of guessing from chat history. You can customize the wording, but the key constraints are:

• →Tool-first: Neotoma is called before composing a response
• →No silent fallback: Claude does not substitute its own memory
• →Grounded output: every fact traces to a tool call in the current turn

Start a new chat inside the project and test with a query like:

show me all my posts

Claude should call Neotoma tools before answering. Look for tool-use indicators (e.g. “Ran commands, used Neotoma integration”) in the response. If Claude answers from memory instead, check that:

• →The chat is inside the project (not a standalone conversation)
• →The connector shows "Connected" in Settings → Connectors
• →Tool permissions are set to "Always allow"
• →The project instructions are saved (not just drafted)

• →Conversation memory (saved memories and chat history that persist across sessions on all plans)
• →Projects (organize chats with scoped documents and custom instructions, 200K context window)
• →Artifacts for generated documents and code
• →MCP server connections via Claude Desktop and claude.ai (remote MCP)
• →The Memory Tool (API beta) for agents built with the Agent SDK, which exposes a client-side /memories file directory for lightweight cross-session state

• ×Structured entity resolution across conversations and projects
• ×Persistent memory that survives session resets and model updates. Claude's memory stores preferences but not structured, schema-bound entities
• ×Cross-tool access; data stays inside Claude's ecosystem
• ×Deterministic state reconstruction from recorded observations
• ×Schema validation or append-only guarantees for the Memory Tool's /memories files. It is a useful scratchpad, but files can be silently overwritten and there is no field-level provenance. See Memory infrastructure for Claude agents for the comparison in detail.

• →Structured entities with canonical IDs that persist across all sessions
• →Deterministic state evolution: same observations always produce the same result
• →Full provenance and audit trail for every stored fact
• →Cross-tool continuity: memory is shared with Claude Code, Cursor, Codex, and ChatGPT

Keep Claude's memory and projects on. They handle conversational context and preferences; Neotoma handles structured state. Both are active simultaneously with no conflict.

• →Memory in Claude (saved memories and chat history)
• →Projects (scoped documents and custom instructions)
• →MCP in Claude Desktop (connecting local MCP servers)
• →Context management (developer platform memory tool)

Once Neotoma is running, try these starter commands in Claude to see cross-session memory in action: