Engraphis 1.3: trust boundaries become explicit

Engraphis 1.3 focused on a difficult part of durable memory: what happens when content is not trustworthy. The release makes ingress policy explicit, keeps provenance attached as memories become graph evidence, and aligns the public front ends around one v2 service boundary.

What changed

• Fail-closed untrusted ingress. Provenance and extractor metadata are allowlisted, and suspicious content is kept from prompt-visible recall until it is safe to use.
• Trust survives derivation. Resolution, structured graph writes, consolidation, and entity profiles preserve the trust state of the evidence they came from.
• Consistent v2 behavior. Public server entry points share the same service, so recall and write behavior do not drift between the dashboard, API, and MCP surfaces.
• Safer deterministic embeddings. Persisted vectors rebuild after an embedding mapping changes, and prompt-facing recall widens its candidate search when untrusted results crowd out trusted context.
• Broader release coverage. CI and release metadata now cover Python 3.13 and 3.14 alongside the existing support lanes.

Why this matters

Memory systems are downstream of everything they ingest. Engraphis 1.3 treats provenance as part of the data model rather than a label that disappears after extraction, ranking, or consolidation.

Try the upgrade

See the full changelog for the security and migration details.

Know where a memory came from.