5.0 GA is shipping today, and we keep shipping. Everything below lands as 5.0.x patches on a continuous train — fast, reliable, no parallel minor branch to wait on. This page is the customer-shaped read on what's already live, what's in the next patch, and what's queued for later 5.0.x patches. Honest framing — no commit dates, no over-claim. We ship when the surface is right. For the briefer card view, see what's next.
Every piece sits in one of three buckets below. Today means it's already shipped in a 5.0.x patch and you can install it now. Next patch means it's landing in 5.0.13 — the design has been settled, code is landing, internal builds carry it. Later in 5.0.x means scoped and named, sequenced for a later 5.0.x patch. The sections after this ledger expand each piece with the customer-side story.
Live in 5.0.x — install now
Landing in 5.0.13
Queued for later 5.0.x patches
Items may slip back or forward between columns as the patch train moves. We update this page when reality shifts. The release notes are the formal cut; known issues tracks 5.0.x rough edges.
Eldric's memory is the spine of how it gets better at your work over time. The 5.0.x patch train sharpens both the boundaries of that memory (who sees what, where it came from) and its feedback loop (what gets reinforced based on what worked).
Each memory entry gains an explicit scope — personal, project, team, tenant — and recall fan-out respects that scope. A teammate searching your shared project namespace sees what's shared; your personal notes stay personal. The dream engine inherits the same scope on the writes it generates from completed sessions.
What changes for you: shared workspaces become safe to use without paranoid pre-sanitisation. The platform enforces the boundary, the customer doesn't.
When the customer accepts an answer, the platform notices. A multi-stage pipeline — acceptance signal → semantic-novelty filter → quality judge → recurrence threshold → classifier update — turns that signal into incremental retraining for the on-device router model. Bad signal is quarantined; the customer's thumbs-up is a vote that lands.
What changes for you: the cluster's intent classification gets quietly better at your queries over weeks, without explicit fine-tuning runs.
The intelligent upload dialog gains a schema-aware ingest substrate — for structured content (tables, CSVs, JSON, XML schemas), the wizard reads the schema before chunking and proposes a strategy that respects record boundaries instead of slicing through them.
What changes for you: ingesting structured data stops requiring per-document overrides. The platform picks the right strategy by reading the data.
The native inference daemon consults associative memory once, at the prompt boundary today. A later 5.0.x patch extends the same hook to fire per generated token — letting the model pull in supporting fragments as it composes the answer, not just before it starts.
What changes for you: longer answers stay grounded throughout, not just in the opening. The trade-off is some additional latency; the gain is fewer mid-answer drifts.
5.0 already ships native inference alongside the optional Ollama path. The 5.0.x patch train promotes the native path to the default on platforms where it's now fast enough, and adds the structured-ML stack on top.
The structured-ML workloads — policy execution, forecasting, vision-language encoding, associative retrieval — have a native compile path. No Python sidecar, no separate model server. The same single daemon serves them directly via a compiled graph. The Apple Silicon and Linux paths land progressively across the 5.0.x train.
What changes for you: faster cold-start, less memory pressure, fewer moving parts to monitor. Especially visible on Apple Silicon, where the native path runs on the unified-memory architecture without GPU-transfer overhead.
For closed-loop control (robotics, industrial automation), a later 5.0.x patch adds a world-model layer between perception and action: ingest sensor / camera frames, predict the next-state, dispatch the policy. Safety gates & wall-clock constraints sit on top so the loop has bounded latency.
What changes for you: structured ML on Eldric isn't just isolated workloads any more. The encoder, forecaster, policy, and retriever compose into a coherent control surface.
The training subsystem ships a round-based federated mode: each site trains locally on data that never leaves the site, the cluster aggregates gradient shards, the next round broadcasts the updated model. Privacy-preserving fine-tuning across a federated deployment.
What changes for you: organisations that cannot pool training data across jurisdictions get a path to a shared model anyway.
Both GGUF and native xLSTM paths get focused performance work in 5.0.13: KV cache reuse improvements, speculative decoding default-on for matched draft models, tighter batching for the streaming token path.
What changes for you: lower time-to-first-token, higher concurrent-session throughput on the same hardware.
5.0 shipped the multi-controller consensus story end-to-end — Raft cluster forms, replicates, fails over under live test. The next patches close the remaining gates that turn it into a hands-off production-HA bootstrap. The deeper story is on the clustering page.
Today, when a controller leader changes, clients reconnect via either an external load balancer or manual reconfiguration. 5.0.13 ships a leader-aware client endpoint shape — the cluster exposes a single connection target, the platform redirects to whichever node is the current leader, the client never sees the failover. The design is settled with edge split-horizon for the admin shell; the implementation is now landing.
A device-fleet inventory at the cluster level: the network plugins, sensors, controllers, and edge appliances the platform manages get a unified registry with capabilities, vendor metadata, and per-device configuration. Plugged into the GUI as a fleet view.
What changes for you: the operations team gets one screen for what the cluster is talking to, not a scatter of per-plugin dashboards.
5.0 shipped the multi-site federation surface end-to-end (federation Raft top-tier, mTLS protocol, cross-branch routing, install wizard). A later 5.0.x patch adds the operational validation work — a multi-region test lab that exercises federation under realistic WAN-latency & partition scenarios, plus a chaos-engineering harness for the HA path.
The final piece of the 5.0 HA story lands as production deployment runs: three-node controller quorum, auth-store unification across leaders, identity replication on recovery. 5.0.13 ships the validated bootstrap path for customers running real cluster control planes.
The chat shell, the macOS GUI, the iOS app, and the CLI all get attention across the 5.0.x patch train — both in shared cross-platform features and in per-platform polish.
The chat shell becomes honest about what the platform can do for this customer, right now. Tools, knowledge bases, native capabilities, plugin-provided extensions — surface inline rather than hiding behind menus. License gates and tenant-enabled features drive the same UI signal.
What changes for you: customers stop asking for things the platform already supports because they couldn't find them.
Per-message visibility badges (will this enter long-term memory? whose memory?), an end-of-output prompt to pin or bookmark, a per-message override menu, and three archetype presets in settings — “remember everything,” “ask each time,” “remember only what I pin.” Asymmetric: the user sees more than the model does about what's being kept.
A first-party batch of platform-native tools: file operations bound to the configured storage, comm operations (mail / SMS / chat protocols), memory operations (store / recall / forget), high-power batches for repetitive work. Available everywhere the chat surface is — web, macOS, iOS, CLI — with consistent semantics.
A sync framework that keeps conversations, prompts, pinned memory, and per-device settings coherent across laptop, phone, and desktop. Two reference domains land first (sessions and prompts); the framework lets later domains plug in without re-engineering each one.
What changes for you: starting a conversation on one device, continuing it on another stops requiring an explicit export step.
5.0 ships the plugin install machinery (catalog browse, signed-archive verify, manifest validate, install / uninstall / update). A later 5.0.x patch adds the customer-side browser UI on top — search, filter, preview, install with one click. Configuration of installed plugins happens in the GUI, not via files.
The macOS GUI and CLI get coverage for the new 5.0.x features — retention UX on both, self-awareness UI on the desktop client, the new native-tool surfaces in the CLI subcommand catalogue.
The platform's extension surface keeps growing — first-party drivers for additional vendors, broader SNMP category coverage, and the long-tail device packs that an operations team accumulates over a deployment's lifetime.
The networking-and-infrastructure plugin build program adds the next batch of Tier-1 platform drivers — server fleet management, enterprise-class platforms with API-documented surfaces. Each one ships as a verified-doc plugin, not against speculative API knowledge.
SNMP coverage broadens — server BMC management (ENTITY-MIB), switch bridge management (BRIDGE-MIB), storage SMI-S, environmental sensors, additional UPS / PDU vendors, surveillance & voice. Each category driver covers a family of vendors via the standard MIB; specific vendor packs layer on top.
For the dozens of niche devices a typical enterprise accumulates, a generic SNMP driver plus a vendor MIB pack architecture: drop the MIB, get a working driver. Operations teams stop maintaining a backlog of “the platform doesn't talk to this one device.”
The per-tenant data-isolation guard expands to cover every module surface, not just the data module. Every endpoint that takes a tenant identifier in the path gets verified against the authenticated tenant, with a clean 403 on cross-tenant attempts. The audit log captures every denial. Already live in 5.0.x; coverage continues to broaden across patches.
None of the pieces above has a release date attached. That's not because we don't have an internal plan — we do. It's because software dates slip; published dates create pressure to ship before the surface is right. Eldric 5.0 took the time it took. Each 5.0.x patch will take the time it takes.
What you can rely on:
dnf update and you're current.The 5.0.x patch train moves on customer signal. If something you need isn't on this page, write to office@eldric.ai. Paid-tier customers with a license ID: support@eldric.ai for prioritised handling — feature requests from active customers carry weight in scheduling.
For the briefer cards view: what's next. For everything shipped right now: release notes. For the deeper clustering story: clustering & HA. For 5.0.x rough edges and preview-status caveats: known issues.