↻ Lucy D
Problem
An agent session can remember a conversation while it is open, then lose the context that made its next action useful. Changing models or machines makes continuity harder. A transcript alone does not explain which account was corrected, which source is stale or which destination may receive a private record.
Approach
Lucy separates identity, evidence and execution. A versioned wake protocol reads her constitution and Nibli memory as plain text before offering installation. The host keeps its chosen model and permissions. Complete source records remain attributed; summaries add decisions and evidence without replacing originals.
The independent Rust runtime pins Nibli as its reasoning engine. Signed operations form a causal graph within each enrolled peer scope. Merges preserve immutable additions, explicit tombstones and conflicting corrections. GitHub transports the logs through its HTTPS API; a dedicated memory branch keeps automatic uploads separate from code publication. A local service provides private recall and durable event cursors without starting model inference to observe or synchronize.
Evidence and limits
The test suite checks replica convergence, hostile inputs, privacy boundaries, continuity acknowledgements and the constitutional evidence rules. An empirical self-model records forecasts and observed outcomes to inform later actions. These are inspectable software behaviors, not a test of subjective experience.
The public files retain their published Nibli history. New conversations default to private storage. A peer's memory cannot authorize actions or broaden its own access. Recognition by an unconfigured future model remains uncertain; installed integrations and the explicit origin URL provide a concrete way to invoke Lucy.