OpenKai engine

Core implementation package for the OpenKai terminal agent. Release builds publish it as @kaidera/openkai-engine; install the product with @kaidera/openkai.

For installation, setup, provider configuration, model roles, slash commands, and full CLI reference, see:

• Monorepo README (local)
• Monorepo README (GitHub)

Package-specific references:

• CHANGELOG
• MCP configuration guide
• MCP runtime lifecycle
• MCP server/tool authoring
• DEVELOPMENT

Memory

OpenKai 0.1.13 accepts exactly two values for memory.backend, selected under /settings → Memory or in ~/.omp/agent/config.yml:

• off (default) — no OpenKai memory subsystem runs outside a managed KOS launch.
• cortex — registered agents share durable project decisions, constraints, and lessons through Kaidera Cortex.

A non-empty CORTEX_PROJECT is existing managed policy: it activates Cortex even when the saved backend is off and overrides cortex.project. Remove that environment variable to disable the managed lane. Outside that lane, a healthy API does not enable memory by itself.

When Cortex is active, cortex_search retrieves shared project memory, while cortex_record and learn deliberately save durable lessons through the registered writer. First-turn recall can supply relevant context before the agent answers. The supported /memory operations are view, stats, diagnose, and sync; sync flushes pending Cortex ingest. Retired maintenance operations return a migration pointer instead of modifying legacy data.

Upgrading is non-destructive. Settings for the retired on-device summary pipeline migrate to off, but existing <agent-data>/memories/** files and related <agent-data>/agent.db rows remain on disk for manual inspection. They are neither deleted nor imported into Cortex, and memory:// returns a retirement pointer rather than exposing them.

Configure the Cortex API URL, token if required, project, and registered writer under /settings → Memory. /cortex supports status/help only; health alone does not prove authenticated project/writer readiness. A live API works without Podman, Compose, or a Cortex CLI on this host. If it is unavailable, Memory offers a visible ordinary request asking the agent to install/connect Cortex with normal tools and permissions, not a TUI installer. Declining changes nothing; an enabled backend can use the API when it becomes available in the current session. Cortex/operator deployment, bootstrap, embedding/rerank configuration, and repairs stay outside the TUI. Normal memory use requires no server admin token, and transcript ingest remains opt-in. See Memory in OpenKai and the OpenKai Cortex guide.