Cordis Memory

cordis-memory is the persistent cognitive store for Cordis. It uses Python's built-in SQLite support: no graph server, vector database, model, or network dependency is required.

It stores five durable kinds of cognition:

• event: a task, plan, tool, test, error, or verification observation;
• episode: a completed task experience;
• knowledge: an evidence-backed fact with provenance;
• pattern: a repeated independently sourced observation with confidence;
• capability and principle: model/tool performance and stable action rules.

Memory scopes are conversation, workflow, project, and global. Project-scoped cognition is never returned to another project. Global knowledge is deliberately shared only when a caller requests global scope.

The relationship graph uses nodes and edges in the same SQLite database. The initial graph records task-to-event and event-to-tool relations; hosts can add their own project, artifact, model, environment, plan, and evidence nodes.

Minimal use

from cordis_memory import CognitiveStore

store = CognitiveStore("cordis-cognition.db")
store.record_event({
    "event_type": "tool_failed",
    "project_id": "my-app",
    "task_id": "task-1",
    "tool": "shell",
    "subject": "npm test",
    "actual": "exit code 1",
})

store.remember(
    kind="knowledge",
    subject="package manager",
    content="This project uses pnpm workspace.",
    scope="project",
    project_id="my-app",
    source_id="repo-readme",
)

snapshot = store.query(intent="package workspace", project_id="my-app")

Patterns begin as candidates and become active only after at least two independent source_id values support the same statement.