The xLSTM family is now a first-class workload type. Admin-side bundle export/import for moving knowledge between installations. Edge-runtime deployment on Raspberry Pi, Intel NUC and NVIDIA Jetson. Full multi-tier memory subsystem with crash-safe persistence. Sixteen science domains. The full multi-worker stack with 4.x feature parity.
Eldric 5.0 ships brain-inspired AI as a server you run on your own hardware — multi-tenant, multi-worker, with portable knowledge bundles, edge deployment from Pi to Jetson, and a structured-ML daemon for the xLSTM family alongside the LLM stack.
The eldric-xlstmd daemon (port 8884) serves four workload classes from one process: policy execution (LRAM), forecasting (TiRex), encoding (ViL), and associative retrieval (Hopfield-style). The router dispatches to it by workload type; the edge gateway proxies; the IoT worker binds policy execution to real-time control loops with safety interlocks. The retrieve backend is native code with microsecond latency on CPU alone. License-gated per workload, structured error responses on missing capability.
Policy-driven control of OPC-UA, Modbus and MQTT-Sparkplug-B endpoints through the IoT worker. Safety interlocks ship by default: magnitude clamp on output, joint-limit checks, watchdog timeout, NaN reject, and an emergency-stop path. Bindings can be pinned to a specific xLSTM worker or set to failover automatically. Streaming via WebSocket transport for sub-tick latency; OPC-UA and MQTT-Sparkplug-B for fleet-wide subscriptions; watchdog-driven fallback (zero / freeze / last-known-good) when the policy worker misses its deadline. Details on xLSTM & IoT transports.
Pack the matrix memory, vector documents, knowledge-base sources, ENRN classifiers, skills, source-registry entries, dream artefacts and identity overlays of one Eldric installation into a single signed file. Ship it. Unpack on another cluster with a clean merge — same version or different. Per-tenant scope filters so an export never carries other tenants' data with it. Useful for tenant portability, project handoff, edge seeding, lightweight federation between regional clusters, and disaster recovery.
Single-node Eldric install on Raspberry Pi 4 (4 GB+ RAM, 32 GB+ storage), Intel NUC and NVIDIA Jetson. Local knowledge-base, local matrix memory, local chat serving without a controller in sight. Buffers writes when the link to the central cluster is down and flushes when it returns. The cognitive layer of an autonomous platform, on the platform.
Multi-tier memory: matrix memory (mLSTM-inspired updates), vector RAG (full-text + semantic search), ENRN classifiers, the Agent × Domain matrix (15 agents × 7 domains). Crash-safe persistence with integrity checksums. Sub-millisecond memory recall on hot paths.
Retrieval-augmented generation ships on by default in 5.0. Upload documents (PDF, DOCX, Markdown, plain text, code, CSV, audio, video, sensor streams) and the platform runs them through content-aware chunking — different strategies per content type, with the intelligent-upload flow auto-suggesting parameters and letting you override before commit. The native GGUF embedding model (~80 MB, runs on CPU) embeds locally; the data worker stores chunks alongside vectors; queries cite the source passages back to the user. Custom classification (Pro+) lets you teach the router your own intent classes. The retention loop turns accepted answers into the next training corpus over time. See using RAG, RAG architecture, chunking strategies, RAG on demand, custom classification.
The Source Registry (§43) groups scientific data sources into sixteen categories with twenty-eight seeded sources and a plugin entry point for adding your own. Eleven LLM-callable tools dispatch through one endpoint. Bioinformatics, pharma, CRISPR, LIMS, materials, climate, neuroscience, particle physics, gravitational waves, space agencies, and more.
Communication worker (email, SMS, VoIP, WhatsApp, Signal, Teams, XMPP). Media worker (STT, TTS, video processing). Training worker (Unsloth, Axolotl, TRL, MLX, llama.cpp, xLSTM backends). Swarm controller, Agent worker, Data worker, Cloud worker, Native Inference daemon — all production-ready with 4.x feature parity.
The macOS GUI client carries a cryptographically-signed in-app updater. New releases are picked up automatically from the package server with a small notification in the app; the update applies on next launch. Update channels are stable by default and beta as an opt-in via the GUI's settings.
A research preview of Modern Hopfield Compressed retrieval ships in 5.0 as an opt-in. Customers can convert a knowledge-base's matrix-memory layer to a compressed representation that's smaller on disk and faster at high concurrency, at a small accuracy trade-off on the kind of edge-case queries where exact retrieval matters most. The full-precision path stays available alongside it. Details on advanced retrieval.
The inference daemon now consults the matrix-memory layer directly at prompt boundaries. Model answers are grounded in your tenant's own data without a separate retrieval round-trip — useful for chat-style workloads where total latency matters and for customer-specific content where the model would otherwise produce a generic first pass. Sub-2 ms per-token overhead on CPU. Pro+ opt-in. Details on smart memory inference.
The iPad client lands as part of the universal-app shape alongside iPhone. Tablet-aware split view (sidebar + chat + artifact), floating composer, Apple Pencil and Scribble, drag-drop ingest from Files / Photos / Mail, notification center for long-running agent tasks. Theme settings sync across all clients via the controller. Multi-window for Stage Manager + Split View ships as preview. TestFlight today, App Store next. Details on iPad.
The "AI that builds AI" flow now has a chat-shell admin interface. Describe the agent you want; the platform generates the agent code, tool definitions, prompts and test cases; sandboxed test runs verify behaviour before deployment. The generator and the builder share the agent package format; either path produces a self-contained agents/<name>/ directory you version-control.
An ENRN v18 router-intent classifier ships as an opt-in for Professional and Enterprise tiers. It replaces the small-LLM routing decision with a single-pass neural classifier trained on your cluster's accumulated routing patterns. Lower latency on the routing decision, less GPU time burned on classification; the LLM stays in the loop for genuinely ambiguous queries. See advanced retrieval for the customer-facing detail.
The minimal-runtime package now ships as a signed aarch64 RPM alongside x86-64. The same one-line installer detects the host architecture and installs the right build. Verified on Raspberry Pi 4, generic ARM64 servers and NVIDIA Jetson. See edge install for the deployment paths.
The IoT worker's xLSTM policy bindings drive WebSocket, Modbus TCP and the publish half of MQTT-Sparkplug-B at GA (Pro+ tier). OPC-UA ships at preview through 5.0 RC1 — the C++ source is in the release, the open62541 library bundles into the cluster RPM at the next cut. The MQTT-subscribe half — pulling observations from a Sparkplug-B sensor fleet through the policy worker — stays preview for the 5.0 line and lands in 5.1. Safety-fallback (zero / freeze / last-known-good) is unconditional across all transports. See xLSTM & IoT and known issues.
Every model in the chat picker shows a small coloured badge identifying the backend that serves it — Ollama, OpenAI, Eldric Inferenced, vLLM, llama.cpp, Anthropic, xAI, Groq and the rest. Green for models on your own cluster, brand-coloured for external APIs. The same badge appears under every assistant message so you always see which backend served it. No silent re-routing between local and external. Customer-facing detail on model providers.
A new Show tool query details toggle in chat settings (default on) prints a one-line summary of what each tool actually queried — knowledge base + namespace + query + hit count for RAG, source + query for science look-ups, search engine + terms for web, recipient for comm. Turn it off for a denser transcript. Available across every chat client.
Agents you build inside the chat shell now have one-click share within your tenant, publish-to-marketplace for the cluster admin to review, and an in-place editor with versioning. Sandboxed test runs validate every change before deploy. The agent-package format stays the same — agents/<name>/ — so manual editing and the UI agree.
Phase D adds Stage Manager + Split View multi-window, a comprehensive keyboard-shortcut map for chat actions, cross-window drag to fork a conversation, and a collision toast when two clients edit the same theme. macOS now also shows long-running tasks as native UNUserNotification entries, with APNs wired through for remote pushes from cluster events. See iPad page.
Admin UI for the four xLSTM workload classes — policy, forecast, encode, retrieve — with per-tenant enable, per-class model selector, and live load chart. Routing requests to the xLSTM worker no longer needs an admin to read JSON.
Existing matrix-memory namespaces can be migrated to the Modern Hopfield Compressed variant on a per-namespace basis with a single command. The chat shell shows migration progress; rollback is a single command. Lets customers adopt the new format on the namespaces that benefit without committing the whole installation at once.
5.0 GA ships signed RPMs for the Fedora / RHEL / CentOS / Rocky / Alma family. Ubuntu 24.04 and Debian 12 DEB packages follow shortly after GA. Customers on Ubuntu / Debian today can run 5.0 via WSL2 or container.
All deferrals are documented on the features catalogue with WIP markers, and the customer-facing list with status sits on known issues.
Customers on the 5.0-alpha line upgrade through the rolling-update orchestrator. sudo dnf update eldric-aios on each node, in the order the orchestrator specifies; the controller drains, installs and restarts each node in turn while the rest of the cluster carries traffic. For single-node installs, the standard dnf update works directly. Tested upgrade path on Fedora 42+, RHEL 9+, CentOS Stream 9+. Customers on 4.x: write to support@eldric.ai for the migration playbook.
The free tier ships with the full kernel: one controller, one router, two workers, all features for evaluation and small-team use. Standard / Professional / Enterprise tiers add capacity, multi-tenant features, the bundle export/import, the xLSTM workload classes, and federated workflows. License files are signed and gate features at runtime — see pricing for the tier breakdown.
To install: get started. To see what changed in detail: what's new. For the full feature catalogue: features. For platform-on-your-domain: apply Eldric to your domain.