MemStrata Client

An open Python API and CLI for a separately installed MemStrata runtime.

This repository contains protocol, validation, transport, examples, and a synthetic mock runtime. It does not contain MemStrata's memory engine, ranking, extraction, query compilation, policy implementation, prompts, evaluation corpora, or licensing enforcement.

Try the open client

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install memstrata-client
memstrata-client init --mock
echo '[{"role":"user","content":"The Atlas migration completed."}]' \
  | memstrata-client process

The mock is deliberately synthetic: it echoes the final message and emits a linked evidence record. It demonstrates the integration and response contract; it does not demonstrate MemStrata's proprietary accuracy or memory behavior.

Python

from memstrata_client import MemStrata

client = MemStrata()  # uses the runtime configured by `memstrata-client init`
result = client.process([
    {"role": "user", "content": "Project Atlas moved to Toronto."},
])

for memory in result.memory:
    print(memory.text, memory.evidence_ids)
print(result.assurance.source_truth)  # always NOT_ASSERTED

The public client validates protocol version, request identity, byte ceilings, unknown fields, duplicate IDs, and every claim-to-evidence reference. Runtime diagnostics are suppressed by default so submitted messages are not copied into client exceptions.

Commercial evaluation runtime

The signed evaluation runtime is a separate commercial artifact. Once issued:

memstrata-client init \
  --runtime /absolute/path/to/memstrata-eval-runtime \
  --license-file /absolute/path/to/license.json
memstrata-client doctor

License contents are sent over the runtime's stdin protocol, not placed in the process command line. The runtime—not this open client—verifies and stores the entitlement.

We do not describe local binaries as tamper-proof. Artifact signing verifies publisher and bytes; compilation raises reverse-engineering cost; neither can stop a privileged owner of a machine from patching code.

Agent frameworks

The core accepts ordinary role/content dictionaries used by OpenAI-compatible histories, AutoGen, and CrewAI. memstrata_client.adapters.from_langchain normalizes LangChain-like message objects without adding LangChain as a dependency.

Evidence verification

Use the independent MemStrata Evidence Kit to verify signed evidence artifacts offline. Integrity does not establish that the source assertion is true; the client preserves source_truth=NOT_ASSERTED.

Security and privacy

• No message telemetry is implemented in this package.
• No MAC address, disk serial, CPU ID, motherboard serial, or OS machine ID is collected.
• Do not pass secrets in message content.
• Report vulnerabilities privately as described in SECURITY.md.

Apache-2.0. See NOTICE.

Distribution and commercial runtime packages

The Python client is prepared for tokenless PyPI Trusted Publishing. The separate commercial runtime can be staged into MSIX, notarized macOS PKG, Snap, AppImage, or Flatpak packages without putting the runtime in this repository. See Distribution and Packaging.

The packaging templates are open build infrastructure. They do not grant a license to, contain, download, or reveal the commercial runtime.