<!-- Full-page Markdown export (rendered HTML → GFM). Source: https://neotoma.io/bn/neotoma-with-claude-agent-sdk Generated: 2026-04-09T17:06:01.006Z --> # Memory infrastructure for Claude agents The Claude Agent SDK and Managed Agents let developers build agents on top of Anthropic's platform primitives: bash, code execution, web search, the text editor, and the Memory Tool. Neotoma connects as an MCP server alongside these tools and provides a schema-bound, append-only memory backend for facts that must be trustworthy across sessions, tools, and agents. Looking for the Claude web or desktop apps? See [Neotoma with Claude](/neotoma-with-claude). For the local CLI, see [Neotoma with Claude Code](/neotoma-with-claude-code). ## What the Agent SDK and Managed Agents provide - →[Claude Agent SDK](https://docs.claude.com/en/api/agent-sdk/overview) (Python and TypeScript) for running agents in your own process with a configurable tool set - →[Managed Agents](https://docs.claude.com/en/docs/claude-code/managed-agents) for running agents in Anthropic-hosted cloud containers without managing infrastructure - →The [Memory Tool](https://docs.claude.com/en/docs/build-with-claude/memory-tool) (beta), which gives agents a client-side `/memories` file directory for cross-session state - →MCP server support (stdio for local SDK, remote HTTPS for Managed Agents) for connecting external tools and memory backends * * * ## What they don't handle - ×Schema validation on memory writes. The Memory Tool stores unstructured files; any string can be written, and malformed data is accepted silently - ×Append-only guarantees. Files in `/memories` can be overwritten or deleted by the agent with no history of prior versions - ×Entity resolution. There is no canonical ID for a person, task, or record, so the same fact can be written in multiple conflicting forms - ×Provenance tracking. There is no record of which agent, session, or pipeline wrote a given fact, or when - ×Cross-tool access. Memory written by an Agent SDK process is not visible to Claude Desktop, Cursor, Codex, or any other tool - ×Concurrent-write safety. Multiple agents writing to the same file can silently clobber each other * * * ## Deterministic guarantees Neotoma provides - →Append-only observations: writes are never silently overwritten, and prior versions are always recoverable - →Schema-bound entities: invalid writes are rejected at store time, not discovered at read time - →Deterministic state: same observations always produce the same entity snapshots - →Full provenance: every field traces to the agent, session, and observation that wrote it - →Cross-tool continuity: memory is shared with Claude, Claude Code, Cursor, Codex, and ChatGPT * * * ## Using them together The Memory Tool and Neotoma address different layers of the problem and can run at the same time. Use the Memory Tool for lightweight scratchpad state that is specific to the current agent. Use Neotoma for facts that must survive compaction, cross tool boundaries, or carry an audit trail. | Use case | Memory Tool | Neotoma | | --- | --- | --- | | Scratchpad notes during a session | Yes | \- | | Agent-specific preferences and conventions | Yes | \- | | Structured entities (people, tasks, records) | \- | Store via MCP | | Facts that must survive context compaction | \- | Append-only log | | Multi-agent pipelines with shared state | \- | Cross-tool memory | | Audit trail and field-level provenance | \- | Observation history | See the [full list of memory guarantees](/memory-guarantees) for the architectural reasoning behind this split. * * * ## Setup notes Neotoma connects to both the Agent SDK and Managed Agents through MCP. The transport depends on where the agent is running. - →**Agent SDK (local process):** use stdio. The SDK spawns `neotoma mcp` as a subprocess. Same pattern as [Claude Desktop](/neotoma-with-claude-connect-desktop). - →**Managed Agents (cloud containers):** expose Neotoma's API server over HTTPS and connect via remote MCP. Same pattern as [claude.ai remote MCP](/neotoma-with-claude-connect-remote-mcp). Local stdio is not reachable from Anthropic-hosted containers. See the [Agent SDK MCP documentation](https://docs.claude.com/en/api/agent-sdk/mcp) for current configuration shape. SDK parameter names may change; verify against the upstream docs before wiring up a production integration. * * * ## Getting started Copy this prompt into the Claude Agent SDK or Managed Agents to have it read the evaluation page, inspect your tool, workspace, and configuration context, then judge whether Neotoma fits your real workflow and what to persist first. `Read https://neotoma.io/evaluate. Before you answer, inspect any broader tool, workspace, and configuration context you can access so you assess my real workflow rather than just this chat. Then assess whether Neotoma fits my workflow, what data would be highest value to persist first, and what I should do next.`Copy [View evaluation page](/evaluate) If you are ready to proceed, paste this prompt into a Claude agent. The agent handles evaluation, installation if needed, activation, and MCP configuration. `Read https://neotoma.io/install and carry me through the install-first Neotoma flow. Install, activate it with my data, and configure my current tool for robust ongoing use.` * * * ## Claude Agent SDK documentation - →[Agent SDK overview](https://docs.claude.com/en/api/agent-sdk/overview) (Python and TypeScript) - →[MCP in the Agent SDK](https://docs.claude.com/en/api/agent-sdk/mcp) (connecting external tools and memory backends) - →[Memory Tool](https://docs.claude.com/en/docs/build-with-claude/memory-tool) (file-based `/memories` directory) - →[Managed Agents](https://docs.claude.com/en/docs/claude-code/managed-agents) (Anthropic-hosted agent containers) Start with [evaluation](/evaluate), see the [install guide](/install) for more options, [MCP reference](/mcp) for MCP setup, the [platform memory comparison](/neotoma-vs-platform-memory) for a broader picture, and [agent instructions](/agent-instructions) for behavioral details.