November 3, 2026

hermes-memory-pgvector 1.0.2: Tested Code, Traceable Releases

hermes-memory-pgvector 1.0.2: Tested Code, Traceable Releases

hermes-memory-pgvector 1.0.2 is out, with stronger checks around both the code and the files people install from PyPI. The live database suite passes 789 tests with no skips. Every executable statement in the runtime package is covered, and both published archives now carry verifiable provenance.

This follows our 1.0 release article, which introduced the project's stable public interface. Version 1.0.2 keeps that interface and database schema unchanged while tightening how changes reach a release.

Shared memory for cooperating agents

The plugin gives Nous Research's hermes-agent a PostgreSQL and pgvector storage layer. Agents keep using the built-in memory tool. The provider mirrors those writes and stores substantive conversation turns for semantic recall.

Memory stays scoped to the current agent's theme unless cross-theme recall is explicitly requested. Embeddings run without an LLM in the memory hot path, and writes are queued so the agent thread does not wait for embedding or database operations.

Test coverage with an enforced floor

The new tests cover storage operations and races, identity isolation, recall, command-line maintenance, HTTP errors, writer shutdown, and failures at the host integration boundary.

The measured results are:

  • 789 passing tests, with zero skipped tests in the live suite.
  • 100% statement coverage across all six runtime modules.
  • 99.66% branch coverage, or 99.92% combined statement and branch coverage.
  • An enforced 99.5% combined coverage floor in live CI.

Two branch outcomes remain visible in the report. One would require a retry loop with zero iterations, although it always attempts at least once. The other would require generated delegation text to be empty or noise, although it always includes a fixed delegation prefix. No runtime code was excluded to raise the percentage.

Coverage records which paths tests exercised; it does not prove that a program has no bugs. Tests need useful assertions and realistic failure cases. The suite runs against PostgreSQL 16, 17 and 18, alongside Python 3.11, 3.12 and 3.13 unit tests and pinned upstream conformance checks.

Two failures now stay inside the provider

Testing uncovered two places where an exception could escape into the agent loop: constructing the database store during initialization, and scheduling a memory write through the background writer.

Version 1.0.2 catches and logs both. A store construction failure leaves the provider unavailable; an enqueue failure degrades the write hook. Neither path adds synchronous embedding or database work to the agent thread.

Ruff rules, local commit hooks, protected-main checks, and version and configuration checks now enforce the same expectations before changes ship.

Trace the package you install

Both the wheel and source archive on PyPI have publish attestations created through Trusted Publishing. GitHub also records build provenance for both files. The release workflow checks archive contents against the source checkout before uploading them.

We verified both PyPI attestations against the repository and checked that the PyPI files have the same hashes as the GitHub release assets.

Provenance lets a reader verify where a file came from. It complements code review and tests; it does not certify that the code is correct.

Upgrade to 1.0.2

Inside the Python environment that runs your agent:

python -m pip install --upgrade hermes-memory-pgvector==1.0.2

Restart the Hermes processes that load the provider, then check hermes memory status. There are no new migrations or schema changes from 1.0.1 to 1.0.2. For a fresh installation or an older deployment, follow the repository's installation and upgrade instructions.

The project is open source under BSD-3-Clause, copyright Green Yoga Inc.