Chronos Nexus sets out to build AI agents that persist and evolve rather than reset between calls. Its organizing conviction is that the language model is interchangeable and the real substance lives in structured long-term memory in PostgreSQL — agents are temporary personas that write "handover reports" so successors inherit context, forming generational lineages. The design layers three tiers: foundations (schema, the SQL gateway, settings), standard services (local model integration with throughput benchmarking, a database-backed task queue, a Telegram bridge, a dashboard), and an advanced cognitive layer for permissions, dialectics, and conflict resolution.
Several design ideas are genuinely distinctive. Social Access Control is a four-tier scheme — private, persona-shared, type-shared, and public-hive — governing what an agent may keep to itself versus share with peers. The Thesis-Antithesis-Synthesis engine stores knowledge as an explicit dialectical process, preserving how a conclusion was reached rather than only the conclusion. A "Grand Jury" mediator persona is meant to resolve conflicting knowledge or permission disputes. Security is enforced by an immutable SQL gateway: a separate component, inaccessible to agents, that intercepts every query to hard-block destructive statements, auto-inject soft-delete filters, and append ownership filters so one agent cannot read another's private data. Agents also use wait-and-resume, suspending their state to the database while awaiting input to free compute.
The codebase backs much of this with real structure: a data model and gateway with a query validator and an injector built on a proper SQL parser, a local-model client and benchmarking modules, agent registry and persona and state code, knowledge-graph and dialectic and access-control modules, a Telegram inbox and outbox bridge, and a dashboard with telemetry and routes. The security gateway and several foundation pieces are the most fleshed-out parts.
Honest context: the implementation is visibly incomplete — its own task tracker shows only a fraction of the planned work marked done, so most advanced features (evolutionary model selection, the full retrieval pipeline, succession) are designed but not yet verified as working. The documentation also carries an unresolved inconsistency about which local-inference backend is used, and there are no real-world results, deployment, or metrics to cite. This entry therefore describes the architecture and design intent — which are substantial and original — rather than claiming proven behavior.