Engraphis 1.7.3: reliable memory workflows, calmer operations

Engraphis 1.7.3 is out. The 1.7 line makes the path from an agent's first turn to durable, inspectable memory more dependable: recall is budget-aware, fact changes are safer, MCP startup is quieter, and the dashboard stays inside clear operating bounds.

The 1.7 line, now hardened

  • Session-aware recall. Smart MCP now reaches relevant context through a budget-binding default, and the Command Code SessionStart integration can deliver bounded project context on a new session.
  • Safer fact changes. Smart writes preserve subject_key and claim_kind, while the correction resolver distinguishes a real value change from a merely similar sentence.
  • Smoother MCP starts. The stdio wire is isolated from library chatter, cached embedding models take a fast local path, and a background warmup removes avoidable cold-start timeouts.
  • Inspectable project workflows. The dashboard adds project-centered memory workflows, paginated history, actionable review state, and Galaxy layouts with bounded physics and deterministic scene identity.
  • Safer operations. Failed resource imports roll back completely, request-scoped service contexts isolate embedded applications, and vector repair work keeps canonical state and derived indexes aligned.

What 1.7.3 adds

This patch release tightens the Galaxy operating contract. Central-field changes now refresh the cached orbit state used by the next fixed slice, dense layouts remain inside their speed cap, and zero-gravity fallback caches are bounded. The scene algorithm identity is updated so cached layouts cannot be confused with the revised placement contract.

Release evidence is part of the same workflow: the scanner version is pinned, the full browser set is installed in release jobs, and the public release stays tied to a known source artifact. These are reliability improvements, not a new capacity or performance claim.

Try the upgrade

$ pip install --upgrade "engraphis[server]" $ engraphis-init --check $ engraphis-init --prefetch # verify the local engine and cache model weights before connecting an agent

Read the v1.7.3 release and the full changelog for the complete changeset.

Durable context should survive the handoff—and leave a trail.

See it running on your machine

> one command, no signup, no cloud