SebaWag

GitHub profile1 plugin in the catalog

Plugins by SebaWag

1 plugins by SebaWag in the catalog, most-starred first.

SebaWagsamskaraRelational memory provider backed by a knowledge graph (MillenniumDB or any SPARQL endpoint): multi-hop recall with provenance on every fact, an ontology that infers subsumption (Incidente ⊑ Evento) at reindex time, per-fact weights that are reinforced by use and decay with disuse, and explicit conflict reporting — when two values disagree, the agent says so instead of guessing. Zero-setup mode: SAMSKARA_ENDPOINT=embedded://seed/demo.ttl serves a built-in demo corpus from an in-process, read-only store (stdlib only, no network, no service to run) for the first minute; it answers the three query shapes the plugin issues and refuses arbitrary SPARQL rather than guessing. Onboarding backfills the graph from the sessions the agent already had (state.db, read-only) and reconciles new entities against existing ones. Disclosure — with session-end extraction enabled (default), the plugin sends the recent turns of a session to the user's own model through ctx.llm (host-owned credentials, no provider override) to extract facts. If the host does not expose ctx.llm — Hermes' in-process memory-provider path does not today — extraction is skipped rather than redirected: the session records its entity mentions lexically and no model is called. The plugin reads no provider API key from the environment and has no default model host. The graph endpoint defaults to loopback and is the only other network call. The CLI tool tools/backfill.py (never imported by the plugin) reads past sessions from state.db and sends them only to the endpoint the user names, via SAMSKARA_LLM_BASE_URL and SAMSKARA_LLM_API_KEY. No telemetry. Credentials (SAMSKARA_PASSWORD) are read through agent.secret_scope.get_secret.Communitymemory