Installing The Memory Adapter

CoreTex has two coordinated Python packages:

Package Purpose
coretex-memory The durable memory engine: event IR, store, validity, M1-M6 execution, retrieval, packing, rendering, and release loading
coretex-memory-agent The agent integration layer: scoped turns, prefetch, session flush, profiles, provider adapters, and a localhost sidecar

The agent package depends on the runtime package. The validator client is a separate application for chain replay and state resolution.

Until ordinary package-index releases are advertised, install the exact pinned wheels published by the live kit:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip

# Download the runtime wheel, agent wheel, and architecture-matched dependencies
# from the URLs and hashes advertised by /coretex/v5/kit/manifest.
python -m pip install --no-index --find-links ./wheelhouse \
  coretex-memory coretex-memory-agent

Verify the wheel hashes before installation. A production installation should then run the supplied state resolver/activator against the live frontier. The activator:

  1. reads the confirmed on-chain state;
  2. fetches the frontier, profile releases, and transition artifacts;
  3. verifies their content hashes and runtime/ABI compatibility;
  4. prepares derived state transactionally;
  5. atomically activates the new release; and
  6. records enough local state for restart, unchanged-state sync, and rollback.

Use the exact sync command shipped in the current public kit. The activator owns all writes to the active release directory and refuses a stale parent, partial download, incompatible runtime, or hash mismatch.

For offline local experimentation, the runtime provides a direct CLI:

coretex-memory init --path ./memory.db
coretex-memory ingest --path ./memory.db --source notes \
  --text "The deployment window begins Friday."
coretex-memory context --path ./memory.db \
  --query "When is the deployment window?" --budget 200
coretex-memory status --path ./memory.db

The CLI also supports correction, retraction, hard deletion, checkpoint, and restore. Its budget is measured by the configured deterministic counter. An application that needs an exact model-token limit should provide that model's tokenizer through the supported counter interface.