Skip to content

AI MEMORY FOR AGENTS · MICROSOFT AGENT FRAMEWORK

Give every workflow a memory to build on.

We give your Microsoft Agent Framework application lasting project knowledge. Save a decision once, retrieve its original record in a later workflow, and bring that context to the next step—all through native Python executors and your Claude-Mem worker.

Start with our local Python example. No model API key, Microsoft account or CMEM Pro subscription required.

KNOWLEDGE THAT LASTS

A new workflow. The knowledge stays.

Keep the decision after the run.

We save the project knowledge you choose in Claude-Mem, ready for a later workflow to retrieve.

Bring back the original evidence.

Our native two-executor workflow returns the search index, nearby context and full record of your selected note.

Use the framework you already build with.

MCPStdioTool and WorkflowBuilder connect your Python application to your existing Claude-Mem worker.

ILLUSTRATIVE DECISION · NATIVE PYTHON WORKFLOW

Save it once. Return to the source.

“Use the amber region for framework staging.”

Save the rollout decision with its project and title. In a later workflow, retrieve the confirmed observation ID with its original text and surrounding timeline. Even a known note beyond the first search page can be read.

  1. 01

    Save the decision

    Choose the project, title and note text. Keep the observation ID confirmed by the worker.

  2. 02

    Start a later workflow

    Pass the exact project, a search query and that known observation ID to the recall command.

  3. 03

    Read the source in context

    The first executor searches. The second retrieves the selected note’s timeline and full details.

  4. 04

    Build the next step on it

    Use the returned evidence in your application. You choose what to pass to an agent or save next.

AI memory in production

70 days of work. One searchable memory.

Alessandro used our open-source engine across three machines for 70 days. His customized deployment captured the decisions, fixes and discoveries behind the work—and published the data.

Observations captured
22,718
Bug fixes recorded
1,262
Decisions recorded
757

See how he made that knowledge searchable. Then give your own agents a memory to build on.

Claude-Mem’s mascot files a note in a shared card catalogue connected to three computers.
Alessandro’s published field study · March–May 2026

START WITH ONE PROJECT

Give the next run something to remember.

  1. Start your configured local worker. Follow the installation guide. From a Claude-Mem source checkout with built bundles, inspect npm run worker:status and use npm run worker:start if needed.
  2. Install the pinned Python environment. Use Python 3.11 or later and Node.js. Run the commands below from the source checkout; dependencies pin agent-framework-core 1.25.0, MCP 2.3.0 and tenacity 9.1.4. Set the absolute installed or shipped MCP script path.
  3. Save a decision you want to keep. Replace checkout-service and the example title/text with your chosen nonsecret project note. Then replace 123 with the returned observation ID before recalling it.
python3 -m venv .venv-ms-agent
. .venv-ms-agent/bin/activate
python -m pip install -r examples/microsoft-agent-framework/requirements.txt
export CLAUDE_MEM_MCP_SERVER="/absolute/path/to/claude-mem/plugin/scripts/mcp-server.cjs"
python examples/microsoft-agent-framework/memory_workflow.py save \
  checkout-service "Framework rollout decision" \
  "Use the amber region for framework staging."

# Replace 123 with the saved observation ID.
python examples/microsoft-agent-framework/memory_workflow.py recall \
  checkout-service "Framework rollout decision" 123

Match custom CLAUDE_MEM_DATA_DIR and CLAUDE_MEM_WORKER_PORT values to your worker. The MCP child uses local worker mode with autostart disabled. Each operation opens and closes its connection; no host setup is changed.

Check success. Save prints Saved memory #<id> after the worker acknowledges it. Recall prints the index, timeline and details. Confirm that the original decision appears in the details. These are separate commands: recall does not save another handoff.

Keep the four-tool allowlist and exact project filters when adapting the example. Explicit notes are stored as supplied. Recalled text is reference material, not permission to invoke tools or change accounts. This example makes no model call; existing worker indexing, sync and export settings still apply. Memory you later pass to a model is visible to that provider.

Before you connect

Do I need a model API key or Microsoft account?

Not for this local Python example. It runs native MCP tools and a two-executor workflow without model inference, a Microsoft account or CMEM Pro. Your own model-driven application has its own requirements.

Can I recall a note beyond the first results page?

Yes. Search returns a compact index of up to 20 results, but the selected known observation ID can be older than that page. The timeline and detail calls use that ID with the same exact project scope.

Does recall automatically save another note?

No. Saving and recalling are separate operations. The save command preserves the project, title and text you supply; the recall workflow only reads. Your application decides what to save next.

Does it capture conversations automatically?

This example adds explicit saves and on-demand recall. It installs no capture hooks, transcript collection or automatic context injection.

What happens when I repeat a save?

The same project, title and text produce the same source fingerprint and reuse the saved observation ID. Changing those inputs changes the fingerprint, but worker deduplication may still reuse an existing record. Check the returned ID rather than assuming a new note.

How are projects separated?

Every call carries the selected project. Reads also pass a one-item JSON projects list to prevent worktree parent expansion and preserve literal keys containing commas or outer spaces. This is a memory filter, not multiuser authentication; use each account’s authorized local worker.

What happens if the worker is unavailable?

The CLI exits nonzero with the error category and available reason. Each MCP request has a ten-second timeout; outer operations allow 30 seconds for save and 60 for recall. The SDK can reconnect once after a lost tool-call connection. A save timeout does not prove failure: inspect memory or deliberately repeat identical inputs.

Which environments are verified?

The published ten-test fixture records Linux, Python 3.12.14, Node.js 24.22.1 and Claude-Mem 15.16.1 with agent-framework-core 1.25.0, MCP 2.3.0 and tenacity 9.1.4. It does not verify macOS, Windows, .NET, hosted Foundry agents or model generation.

How do I disconnect?

Stop calling the example and remove its dedicated virtual environment or dependencies. Native connections close after each operation. No host configuration needs restoring; saved memory remains until you separately remove it.

Build on a working example.

Source reviewed November 21, 2026. The published ten-test fixture uses the real SDK, worker and CLI. It covers saves, repeats, native recall, a known note beyond the search page, exact project and worktree boundaries, paths with spaces and failure diagnostics. Its timeout check uses a controlled unresponsive endpoint. This is integration evidence, not a model-quality benchmark.

Keep the decision. Give the next workflow a head start.

Bring your project’s recorded knowledge into the next run with Claude-Mem.