AI MEMORY FOR AGENTS · CREWAI
Give your next CrewAI flow a head start.
We keep the decisions your workflow needs beyond a single run. Save a handoff through CrewAI’s native Flow, then find its context and retrieve the original record when the next step needs it.
Explicit save and recall through your local worker. This standalone Flow needs no model key or CMEM Pro account.

KNOWLEDGE THAT LASTS
Carry the decision into the next step.
Hand off knowledge, not another explanation.
We save the note your workflow supplies so the next step can use the decision you already made.
Find the original after more work arrives.
A confirmed save ID keeps your handoff retrievable even when newer notes move it out of the recent index.
Keep each project’s context together.
Our example applies the caller’s selected project to saves, search, timeline and record retrieval.
ILLUSTRATIVE HANDOFF · ACTUAL FLOW SEQUENCE
A decision today. A reference for the next run.
“Keep the scheduler export in UTC.”
The native Flow carries the confirmed save ID into recall. It retrieves the intended note instead of assuming the newest search result is the answer.
- 01
Save the decision
Supply the handoff note, project and logical run ID. The native Flow saves that note.
- 02
Keep the receipt
The worker confirms the saved record’s ID before recall begins.
- 03
Find project context
The next Flow step reads the recent project index and the saved record’s timeline.
- 04
Read the original
Retrieve the confirmed record even when newer notes move it beyond the recent index.
START WITH ONE PROJECT
Save one handoff. Read it back.
- Install Claude-Mem and start the local worker. Follow the installation guide. From a source checkout, check
npm run worker:statusand usenpm run worker:startif it is stopped. - Use Python 3.11 or later and Node.js. Run the commands below from a Claude-Mem source checkout to create a dedicated environment with CrewAI 1.15.39 and MCP 1.30.0.
- Choose the MCP script, project and run ID. Point to your trusted installed or shipped MCP bundle. Match custom
CLAUDE_MEM_DATA_DIRandCLAUDE_MEM_WORKER_PORTvalues to your worker. Save a note you intend to retain.
python3 -m venv .venv-crewai-memory
.venv-crewai-memory/bin/python -m pip install -r examples/crewai/requirements.txt
export CLAUDE_MEM_MCP_SERVER="/absolute/path/to/claude-mem/plugin/scripts/mcp-server.cjs"
.venv-crewai-memory/bin/python examples/crewai/memory_workflow.py \
--project shop-api \
--run-id scheduler-handoff-1 \
--note "Keep the scheduler export in UTC."Check success. Output includes saved_id, recalled and CrewAI’s flow_id. Confirm that the recalled text contains your note. Repeating the same project, run ID and note returns the original saved ID.
Keep one native MCP client open while the Flow runs. Preserve call_tool_result() error checking and the explicit project list when adapting it. If embedding in an app, apply the example’s telemetry and tracing choices before importing CrewAI.
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.

Before you connect
Does this record every CrewAI conversation?
No. This example saves only the handoff supplied by the caller. It installs no transcript hooks or startup injection and does not use CrewAI’s separate memory or embedding subsystem.
Do I need a model key, CrewAI cloud or Pro?
Not for this deterministic local Flow. It needs a running local Claude-Mem worker, Python and Node.js. Model-driven applications have their own provider requirements. AMP/cloud and remote authenticated MCP were not verified by this example.
How do repeated runs work?
The same project, run ID and note reuse the stored receipt. Each invocation has a new native Flow UUID; a duplicate observation retains its original provenance. Changing the note creates a new explicit save. These IDs do not resume a past conversation.
How are projects kept separate?
The helper fixes the caller’s project and sends an explicit one-item project list for reads. Published tests cover worktrees and literal project keys. Use a trusted local worker: these filters are not account authorization.
What happens if the worker fails or times out?
The example makes one attempt and reports failure; it does not automatically retry a write. A failed recall does not undo a confirmed save. Check your worker and saved memory before deliberately repeating the same project, run ID and note.
What has been verified?
The published Linux checks use CrewAI 1.15.39, MCP 1.30.0, Python 3.12.14 and Claude-Mem 15.17.0. They exercise the actual Flow, native client and worker without a model call. macOS, Windows, cloud deployment and a model choosing tools are not established by that fixture.
What data is retained, and how do I disconnect?
The script sends the explicit note, project, IDs and provenance into the local memory path and prints the final recall. No additional model or memory service receives them from this script; your worker’s existing sync and export settings still apply. Stop using the example to disconnect. Its client closes the MCP child; removing the environment does not delete saved memory.
Build on a working example.
Source reviewed November 20, 2026. The published nine-check fixture covers native save/recall, repeated receipts after newer notes, project and worktree boundaries, rejected calls, deadlines and telemetry settings. It exercises the real client and worker without a model call; it is not a benchmark of model quality.
Keep the decision. Carry the work forward.
Give the next CrewAI flow a useful starting point with memory it can trace back to the original note.