Glossary.

Worker a daemon that does the work

A long-running process that does a specific job under the Controller's orchestration — ingests sensor data, runs inference, stores vectors, delivers messages, runs training. "Worker" is the role; "daemon" is the process shape. Same thing, different angle.

In 4.x each Worker is its own systemd unit (eldric-workerd, eldric-datad, eldric-agentd, …). In 5.0 they're role modules hosted in the single eldric-aios binary. The filename convention -d suffix is the classic Unix "I'm a daemon" marker.

Daemon

A long-running background process under systemd. In Eldric all Workers are daemons; the Controller, the Router, and the Edge are also daemons. We prefer "Worker" when talking about the role the process plays, "daemon" when talking about its process lifecycle.

Module 5.0 only

A role host compiled into eldric-aios. Fifteen of them: controller, router, edge, cloud, inference, data, agent, swarm, media, comm, science, training, iot, nova, dream. Each owns a port range and an API surface. The module is what a 4.x Worker becomes when you collapse everything into one binary.

Controller

The orchestrator. Tracks which workers are online, what models they serve, what the cluster topology looks like. Every worker registers with the controller at start; every worker re-registers on heartbeat. The controller itself is a Worker (role: controller).

Router

The load balancer in front of the workers. Clients hit one router; the router picks which worker serves the request based on model availability, worker load, and the configured strategy (round-robin, least-connections, least-latency, AI-decided). Also a Worker (role: router).

Edge

The external-facing gateway. Terminates TLS, enforces API-key auth, applies rate limits, proxies to the router, hosts the chat UI at /chat, runs the plugin system. Edge is where chat.eldric.ai lands.

Extension Tier 2, loadable

The only runtime-loadable thing in Eldric. A specific named capability that plugs into a role host — an OPC-UA connector, a NASA API adapter, a CUDA backend, a vLLM wrapper. Ships as a signed .eldext bundle, runs as a subprocess (Model A) or external service (Model B) under the host's supervision. Extensions expose tools the Brain can invoke.

Plugin chat-shell concept

A toggle in the chat sidebar. Rendering add-ons (markdown, code, charts), data sources (one per knowledge base / API), project scopes, workgroup sharing, privacy opt-outs, tools. Every loaded Extension with category: data or category: tool auto-surfaces as a Plugin. When users enable multiple data Plugins, one chat query fans out to all of them in parallel.

Data-plugin fan-out

The retrieval pattern: when a chat request lands, Edge queries every enabled data.* Plugin in parallel via the retrieval.<plugin-id> contract. Returned snippets merge into a synthesised system message prepended to the conversation before the LLM sees it. The fan-out doesn't care whether each plugin answered via vector search, tree reasoning (PageIndex), live sensor query, or anything else — same plumbing, different epistemic strategies.

Matrix Memory

The data role's compressed associative memory — outer-product updates (M = decay·M + importance·(v⊗k)), matrix-vector recall. Captures patterns rather than exact documents. Complements vector storage (which handles exact retrieval). Hierarchical: Domain → Project → Run matrices cascade from general knowledge to specific context.

Dream

The background consolidation scheduler. During idle periods, walks the cycle ingest → extract → probe → distill → checkpoint → complete to reorganise what's been learned — sleep, but for AI memory. Per-user scope via the sharing.dream plugin opt-in.

Brain

The inference side of things — the LLM(s) that answer queries. Learns new tools at runtime from ETP tool manifests (no retraining). Talks to Extensions via the Eldric Tool Protocol over gRPC / HTTP.

ETP — Eldric Tool Protocol

The wire protocol between the Brain + role hosts and Extensions. gRPC over Unix domain socket (canonical), HTTP/JSON (Model B fallback). Semver-stable for ≥5 years from 5.0 GA. Not MCP — MCP is an optional Phase-4 bidirectional gateway for external clients.

Controller / Worker / Edge / Router — who talks to whom

External client → Edge → Router → Worker(s) → Extension(s). The Controller sits out-of-band, tracking who's online. In a single-node install all five are the same eldric-aios process on port 8880.

ENRN — Eldric Native Neural Network 5.0 shipping, expanding in 5.0.x

A purpose-built native neural-network format used inside Eldric for classifiers, routers, and small inference heads. Trained server-side on the data the user has explicitly chosen to keep, distilled into a compact runtime artefact, and hot-swapped without restart. The continuous-learning loop (next 5.0.x patch) turns customer interaction into training pairs (with opt-in plus a verifier gate) so the model on disk keeps catching up to the workload running through it.

EMM — Eldric Matrix Memory 5.0 file format

The on-disk binary file format for the compressed associative memory described under Matrix Memory. Versioned binary container with a header, blocks of matrix data, integrity hashes, and a write-ahead journal. The .emm extension is what you find in memory/ on a running cluster — one file per matrix per scope (domain / project / run).

mHC — matrix Hopfield Consult 5.0 hook, per-token in 5.0.x

The mechanism by which the inference daemon consults the data role's matrix memory during a generation. At 5.0 GA it fires once at the prompt boundary; a later 5.0.x patch lands the per-token hook that fires inside the decode loop, letting the model condition on what the cluster has learned without expanding the prompt. The customer-facing name is smart memory inference; mHC is the internal term.

L1 native — native xLSTM backend preview in upcoming 5.0.x

A native inference path for xLSTM-family models that runs without llama.cpp, vLLM, or any external runtime. 5.0 ships the scaffold (off by default); a later 5.0.x patch flips it on as the default native backend, with byte-exact roundtrip parity verified against the reference implementation. Same kernel, same API surface — just a faster substrate underneath.

World model in development

Forecast, policy, and perception capability for closed-loop applications — robotics, autonomous systems, industrial control — in development for Eldric's structured-ML / xLSTM tracks. The intent: predict forward in time so a policy can reason about consequences. Scoped to the IoT, robotics, and datacenter tracks, not a general-purpose world-model service. Approach and availability are still being finalised.

Sovereign AI compute

The class of workload run by national AI factories, member-state-funded compute operators, and public-research HPC fabrics. The audience that needs jurisdiction-pinned data residency, sovereign procurement, and security frameworks the hyperscaler can't deliver. Eldric is the AI-datacenter OS that this audience runs.

AI factory customer-shape, generic

A facility that operates AI compute capacity at scale — typically a member-state-funded or consortium-led programme co-located with supercomputing centers. Eldric uses the term as a generic customer-shape, lowercase, recognize-without-naming any specific programme. The infrastructure pattern is what matters; programme names change with politics.

Honest ledger framing discipline

The convention Eldric uses on public copy for any not-fully-live feature. Three columns: today (live, verifiable in the running cluster) / in flight (the next release cycle) / on the roadmap (future). The discipline rule: claim what's GREEN, name what's pending, give an honest gate-status. Customers do not hate a roadmap — they hate surprises.

Gated visibility memory-scoping, next 5.0.x patch

The default-private write + scoped recall pattern in the memory-scoping work landing in the next 5.0.x patch. Memory writes land at the most restrictive scope by default (you-only, this-conversation-only, this-tenant-only) and only widen with explicit opt-in. Recall fans out across visible scopes; never leaks across tenants without an audited promotion. Customers see this as "your conversation memory stays yours, unless you choose to share it."

Miss a term? Tell office@eldric.ai.