Engraphis 1.5: stronger memory quality gates

Engraphis 1.5 is the current GA release: a security and reliability milestone for the v2 memory engine. It advances the schema to 11 and makes the system more deliberate about what can enter recall, how embeddings are kept coherent, and how releases are verified.

What changed

• Governed recall recovery. Existing memories that need review are migrated to explicit approval states, while pending and quarantined content stays out of prompt-visible recall and derived graph evidence.
• Embedding-space safety. The active embedding-space fingerprint is now durable. Space changes trigger a rebuild, and vector recall stays disabled while an interrupted or mixed-space rebuild is being repaired.
• Model provenance controls. Remote embedding models, rerankers, and chunk tokenizers can opt into immutable Hugging Face revision checks, with remote code execution disabled by the loader.
• Operational review tools. The engraphis-cli review list|approve commands and new health diagnostics make review state and prompt eligibility inspectable without editing the database by hand.
• Reproducible release evidence. The release gate now covers pinned evaluation fixtures, type checks, distribution-content verification, SBOM generation, repeat-build comparison, clean wheel and source installs, entrypoint smoke tests, and CodeQL evidence.

Why this matters

Durable memory needs more than retrieval quality. It needs clear boundaries around untrusted input, a recoverable path when the index changes, and evidence that the artifact being installed is the one that was tested. Engraphis 1.5 makes those guarantees visible in the engine, the CLI, and the release process.

Try the upgrade

Read the v1.5 release for the headline changes, or see the full changelog for the detailed security and migration notes. The local core remains Apache-2.0 and local-first; hosted workflows are optional.

Trust the memory. Verify the release.