Local memory. Recorded lineage.
Clear limits.
MADB stores agent memory on your machine and lets you follow recorded relationships. Use this checklist to decide whether it fits your application.
What the terms mean
Memory: information stored beyond one conversation. Lineage: links supplied between recorded inputs, decisions and outcomes. Admission: a decision about whether proposed content enters memory. Cortex: an illustration of this memory architecture.
Choose it for the right job
Consider an evaluation if you need local durable memory and inspectable recorded links. The evidence here does not establish compliance certification, complete activity capture, authenticated delegation, protection against every injection or proof of internal model reasoning.
Release and evidence contract
| Artifact | What to expect | What is not established |
|---|---|---|
| 0.2.14 deprecated release | Release notes and machine-specific benchmarks | Universal performance or client compatibility |
| 0.2.15 public release | Engine and MCP available on PyPI | A supported production security service |
| 0.2.15rc4 fixture | One synthetic run; three durable records; one normal restart; 14 defined checks | External Claude/GPT trials, large-scale deployment or live incident prevention |
| 0.2.15rc2 supplied reel | Recording of a spoofed-payment scenario and claimed policy block | Independent reproduction of its timing and protection claims |
Verify in your environment
- Record your OS, Python version, client version and installed MADB package versions before testing. Unpinned install commands resolve the available package, not necessarily the fixture build.
- Register the server with your client. Confirm the memory tools appear and the intended skill is installed; automatic use is host-dependent.
- Store a harmless test fact and obtain its actual event ID. Use that ID for a second linked record; do not fabricate a parent ID.
- Start a fresh session. Inspect actual recall and trace tool responses; check an unrelated query too.
- Try a change and a deletion on test data. Check historical traces and retained metadata rather than assuming physical erasure.
- Record failures and remove the test integration if its behavior does not meet your needs.
Compatibility status
The 0.2.15 release is qualified on macOS ARM64 and Linux x86_64/ARM64, with MCP Python 3.10–3.13. The maintainer reports 12/12 clean public-install checks and actual CLI-host acceptance for Codex and Claude Code. Windows is deferred; public plugin-store listings are not live. Desktop-UI acceptance is not claimed. The older rc4 fixture still has zero external model trials and remains separate evidence.
Privacy, license and removal
Review telemetry and retention, deletion limits and local use versus commercial integrations. The public MCP package identifies a proprietary license; embedding and hosted-service rights need separate terms. Local storage does not prevent your model host from receiving retrieved context.
To stop using the integration, disable or remove its MCP registration in your client and restart the client. This stops tool access; it does not delete stored files. Back up anything you need and review retention before removing the configured data directory. Exact client removal commands and export coverage still require version-specific verification.