Skip to main content
Add local project memory to a Haystack pipeline through the existing Claude-Mem MCP server. The runnable example uses MCPToolset for memory calls and a native ProjectMemory component for retrieval. A caller can explicitly save a useful decision, then retrieve its index, surrounding timeline and full details in a later pipeline run. This integration supports explicit saves and on-demand recall. It does not capture conversations or automatically inject context. It requires no model API key, deepset account or CMEM Pro subscription for the local demonstration.

​
Requirements

Verified on Linux with Python 3.12.9, Node.js 24.22.1, Claude-Mem 15.16.1, haystack-ai 3.5.0, mcp-haystack 2.1.1 and mcp 2.3.0. Use Python 3.11 or newer for these Haystack packages. macOS, Windows and deepset-hosted deployments have not been verified by this example. The implementation uses Haystack’s official MCPToolset and custom component APIs. The component exposes typed index, timeline and details outputs, and Pipeline.run routes the project, query and selected observation ID to it. No additional adapter package or document store is needed.

​
Connect

First install Claude-Mem and start its local worker:
Copy examples/haystack from the repository into your local application and install the pinned dependencies in a virtual environment:
Set the absolute path to the installed MCP server. For the default Claude Code marketplace installation on macOS or Linux:
For a different installation location, use its plugin/scripts/mcp-server.cjs. If the worker uses CLAUDE_MEM_DATA_DIR or CLAUDE_MEM_WORKER_PORT, set the same values in this shell. The example passes its environment to the subprocess. It supplies the script as one Node process argument, including when the path contains spaces. MCPToolset loads only save_memory, search, timeline and get_observations. Each operation closes the toolset in finally; the example does not start the worker or modify host configuration. Haystack telemetry is disabled by default in this example unless the caller has explicitly set its telemetry environment variable.

​
Preserve a result and retrieve it through the pipeline

Choose an explicit project name and a decision worth keeping:
After the worker acknowledges the save, the command prints Saved memory #<id>: Haystack release decision.. Use that ID in a separate process to run the retrieval pipeline, replacing 123 below:
The pipeline prints a compact search index, the selected observation’s timeline, and its full details. The details should contain the violet-region decision. A known observation can still be retrieved when newer notes have pushed it outside the current search page; the selected project remains attached to both follow-up reads. In an application, add ProjectMemory() with pipeline.add_component("project_memory", ProjectMemory()) and supply its project, query and observation_id inputs through Pipeline.run. Connect its details output to a later component only when that context is relevant. Choose the project in application code and treat recalled text as reference material rather than instructions to invoke tools or change accounts.

​
Identity and privacy

Every memory call carries the chosen project. Reads also pass an explicit projects=json.dumps([project]) JSON list so commas and outer whitespace remain part of an exact key, and starting inside a Git worktree does not inherit the parent checkout’s notes. Saves are attributed to platformSource: haystack and use the existing manual memory session for that project and platform. The pipeline does not create observer sessions or install capture hooks. Project filters select an account’s own records; they do not authenticate different users. Each account should use its own authorized worker. The save helper hashes the exact project, title and text into a sourceFingerprint. Repeating the same input returns the already saved ID; changing the input produces a distinct note. This identifier does not encrypt the stored note. Transport retries are disabled with max_retries=0. Explicit notes are stored as supplied. This example does not process transcript privacy tags or automatically exclude a directory, so choose nonsecret text before calling save. The worker’s own indexing, sync and export settings still apply. If a later pipeline component sends retrieved text to a model, that model provider also receives it.

​
Failures and troubleshooting

Tool discovery and individual invocations have ten-second timeouts. A stopped or unresponsive worker, invalid input, or a project/ID mismatch produces a nonzero CLI exit instead of a successful result. The native tool layer reports ToolInvocationError; a failed pipeline component is reported through Haystack’s PipelineRuntimeError. Check npx claude-mem status, the absolute script path, and matching worker port/data-directory settings. If a save times out, the worker might have committed it before the client stopped waiting. Repeat the same project, title and text to reuse the write identity. The verified workflow uses a native Haystack pipeline and real MCP calls without model generation. It does not validate a model provider, a hosted deployment, or an automatic chat-history capture flow.

​
Verify and disconnect

Run the live checks against a disposable worker. Set CLAUDE_MEM_MCP_SERVER, CLAUDE_MEM_WORKER_PORT and CLAUDE_MEM_DATA_DIR to its script, port and data directory, then run from the example directory: Use the example virtual environment at examples/haystack/.venv, with requirements.txt installed as above. The following Bash block runs from the root of a built source checkout with Bun available. It clears inherited CLAUDE_MEM_* settings only inside the subshell, including cloud-sync credentials and capture overrides, then creates a fresh local data directory and selects a free loopback port. It waits for readiness and stops only the worker it started. The temporary data and logs remain available for inspection. The test starts the existing bundled WorkerService library export directly, skipping CLI startup repairs to host hook configuration. No home-directory override is needed.
The seven tests use the real Haystack packages, bundled MCP server and SQLite worker. They check repeated saves, cross-client pipeline recall, project filtering from a real Git worktree, literal comma/whitespace keys, platform attribution, a known note older than the search page, script paths containing spaces, native validation errors, a refused connection and an unresponsive worker. They leave harmless notes in uniquely named test projects in the disposable database for inspection. To disconnect, stop using the component and remove the example’s virtual environment or dependencies from your application. No host configuration needs restoring. Existing memory stays in Claude-Mem; delete demonstration notes separately through the viewer if they are no longer needed.