AI MEMORY FOR AGENTS · SMOLAGENTS
Give your next agent a stronger starting point.
We turn the project knowledge you choose into memory your SmolAgents application can retrieve. Save a useful decision, bring back its original record, and give your next task evidence to build on.
Start with our local release-checklist demo. No model account or CMEM Pro subscription required for this verification step.

KNOWLEDGE THAT LASTS
Keep the knowledge. Start further ahead.
Keep useful knowledge within reach.
We save selected project notes in Claude-Mem so later work can retrieve them.
Bring back the evidence.
The example checks the confirmed record before reporting successful recall.
Keep your application in control.
Four focused memory tools bind requests to the project you choose.
A REAL NOTE · AN EXPLICIT MEMORY WORKFLOW
A release checklist that survives the run.
“Run the smoke suite before releasing the application.”
Our model-free demonstration saves this checklist and verifies its original record. A model-driven application can then choose how to use recalled evidence in its next task.
- 01
Save the checklist
Store a useful note in your selected project.
- 02
Keep the confirmed ID
The save receipt identifies the original record.
- 03
Read the original note
Retrieve the saved details, even beyond the first search page.
- 04
Use it in the next task
Your application chooses how recalled evidence reaches its agent.
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.

START WITH ONE PROJECT
Save a checklist. Verify you can find it again.
- Start your configured local worker. Follow the installation guide. From a Claude-Mem source checkout with built bundles, inspect
npm run worker:statusand usenpm run worker:startif needed. - Install the pinned Python environment. Use Python 3.10 or later and Node.js. From the source checkout, install the pinned SmolAgents 1.26.0, MCPAdapt 0.1.20, MCP 1.30.0 and websockets 16.0 environment. Set the absolute installed or built MCP script path.
- Save the demonstration checklist. Choose your project key and its absolute working-directory path. The CLI saves the fixed checklist shown above and verifies that its record can be retrieved.
python3 -m venv .venv-smolagents
.venv-smolagents/bin/python -m pip install -r examples/smolagents/requirements.txt
export CLAUDE_MEM_MCP_SERVER="/absolute/path/to/claude-mem/plugin/scripts/mcp-server.cjs"
.venv-smolagents/bin/python examples/smolagents/memory.py \
--project "my-application" --cwd "/absolute/path/to/my-application"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. The memory tools share a connection inside their context manager, which closes when the workflow ends; no host setup is changed.
Check success. The JSON output includes your project, a confirmed observation_id, recalled: true and model_requests: 0. Your ID can differ. Repeat the same command to retrieve the existing checklist identity; an older record need not appear in the recent index.
Keep the memory_tools context open and pass its tools to ToolCallingAgent with your chosen model and add_base_tools=False. The full guide includes this application example. Recalled text is reference material, not permission to run actions. The CLI makes no model call; existing worker sync, export and notification settings still apply. Memory you later pass to a model is visible to that provider.
Before you connect
Do I need a model account?
The fixed local demo makes no model request. Your model-driven agent uses its configured provider and credentials.
Can I save my own notes?
Yes. In your application call save_memory through the connected tools with the text and title you choose. The CLI demo saves a fixed release checklist.
Why use pinned dependencies?
The verified adapter uses MCP 1.x transport names and its websocket import. Keep the isolated pinned environment; do not mix it with an MCP 2.x example.
What happens when I repeat the command?
The same project and trimmed title/text produce the same fingerprint and reuse the saved identity. Retrieval uses the confirmed ID even outside the first search page. Changing content changes the fingerprint, but worker deduplication may still reuse a record.
How is project scope enforced?
The adapter binds all calls and JSON-encodes the selected read scope. Project and source metadata are not agent inputs. These filters are not account authorization; use appropriate worker and data separation for different trust boundaries.
What happens if the worker fails?
Errors fail before success is accepted. Requests default to 30 seconds; connection startup and cleanup have a 65-second watchdog budget. This is not a deadline for an entire agent run. No automatic retry is added; inspect uncertain writes before explicitly retrying.
What gets captured?
The selected notes supplied by your application. This example adds no automatic transcript capture or startup injection. Existing worker sync, export and notification settings still apply.
What has been verified?
The published guide records SmolAgents 1.26.0, MCPAdapt 0.1.20, MCP 1.30.0, websockets 16.0 and Claude-Mem 15.16.1 with Python 3.12 on Linux. Native fixture tests and a separate disposable-worker verifier are documented. Model-generated answers, macOS, Windows and hosted authentication are not established by that evidence.
How do I disconnect?
Remove the memory tools from your agent and close the context. Delete the isolated environment if no longer used. Stored notes remain; no host MCP configuration or provider settings need to be restored.
Build on a working example.
Source reviewed November 22, 2026. Ten native-fixture tests cover scoped tools, fingerprints, errors, timeouts and connection cleanup using a deterministic model. A separate real-worker verifier checks persistence, repeat saves after newer notes, project and worktree isolation, and paths with spaces. These are integration checks, not model-quality benchmarks.
Keep the knowledge. Build on it again.
Bring your project’s recorded knowledge into the next run with Claude-Mem.