Hiveram
Asking an AI assistant about Hiveram? Give it hiveram.com/docs/SKILL.md.
Agent platforms and tracing tools already record traces, reviews and approvals. If yours lets someone outside the work follow one item from the request, through the evidence, to who accepted it, you may not need Hiveram.
Where that chain breaks across tools, the record connects it: the work order, the evidence by reference and hash, the review and the closure, written through the API, MCP server or CLI, beside whatever runtime you use. Hiveram does not run your agents and holds no credentials to your other services.
The sections below describe the mechanics: the work graph, authority modes, memory across a fleet, portable reasoning, payload sizes, complexity tiers, what ships, and how Hiveram sits beside NeuroRouter. For what the record means to the people who answer to an auditor or a client, see the homepage and Why Hiveram.
Underneath, Hiveram keeps a canonical work graph: it classifies every work order C1-C4, routes work to the cheapest capable execution tier, and preserves the reasoning that still matters when work moves across agents, sessions, machines, or disconnected environments.
The same task can cost many times more when it lands on the wrong execution tier, and again when the next agent has to rediscover the last good understanding from chat history. Hiveram turns both problems into product surfaces: routing policy, explicit authority modes, bounded portable reasoning handoff, and visible scaffold review before continuation.
If your team also uses NeuroRouter, Hiveram fits into that wider stack. If you do not, Hiveram still stands on its own as the work graph and handoff layer. Hiveram makes AI work portable without making authority ambiguous.
Git shows where code came from. Hiveram shows where work came from, where it lives, and where it is going.
One connected system of record. This is the normal mode for canonical writes, imports, closure evidence, and team coordination. Everyone points at the same ledger before trusting a result.
Intentional offline or disconnected work. Reasoning can move as a bounded artifact, results can come back later, and final application still happens against a reviewed canonical ledger instead of silent background merge magic.
When a person or agent writes memory, Hiveram records the identity key that wrote it and creates a new revision. You can inspect the history, take an edit lease so concurrent agents do not overwrite one another, work in a draft, and promote it when ready. When work diverges, memory can branch and merge without losing the canonical record.
The memory stays in your Hiveram ledger, whether you use a managed deployment or run it in your own environment.
Portable reasoning is the product answer to long sessions, agent specialization, and airgapped environments. Instead of emailing a giant conversation, teams move a bounded bundle: the work order, its live constraints, key decisions, relevant evidence, and a machine-readable mission briefing for the next agent. A senior agent can architect once, then many cheaper or specialized agents can execute from the same bounded record without rediscovering the project.
Fresh agents start from the current work contract: objective, constraints, decisions, rejected paths, evidence, dependencies, and acceptance criteria. They do not need the whole chat, and they do not have to guess what still matters.
The minimum contract for a short execution pass: immediate objective, targets, files, and reporting expectations.
The bounded reasoning that still matters: decisions, constraints, rejected approaches, dependencies, and evidence.
Frozen full-fidelity context that can be preserved, transferred, and recalled later without forcing it into every live model window.
Every work order carries a complexity tier, model hint, cost estimate, and warning when the claimed model does not fit the task. Provider names can change. The routing logic stays the same.
| Tier | Work shape | Model fit | Cost guardrail |
|---|---|---|---|
| C1 Mechanical |
Fixes, docs, typos, lint, test runs, variable renames. | Small fast model or CLI. | premium-never. Keep routine work off premium reasoning tiers. |
| C2 Implementation |
Implement, add, port, extend, wire, or test known patterns. | Default delivery model or CLI. | Escalate only when execution evidence says so. |
| C3 Judgment |
Design, debug, optimize, security, and work with explicit risks. | Strong reasoning model with review. | Spend more only when underpowered execution is the real risk. |
| C4 Architecture |
Architecture, reframing, pre-mortems, novel problems, vision. | The current frontier model. | Reserve top-tier spend for work that benefits from it. |
One problem, one record. Teams stop multiplying duplicate specs when several agents notice the same thing.
An agent starting work needs the current contract, not the entire record. A briefing resolves open work, recent decisions, and active constraints into a bounded packet for that context window. The agent starts warm; the durable history stays in the ledger.
Export a focused work package, email or file-copy it somewhere else, and get back a reply that still carries who asked, what was done and the evidence by reference.
See the current tip, the recent path that supports it, and which claims still bear weight before a fresh session continues. Teams stop restarting from transcript folklore.
See what flew out, what came back, what branched, and what was applied. Handoff becomes auditable instead of folkloric.
Disconnected teams can keep reading the shared backlog, queue delayed writes intentionally, and apply them later instead of pretending local state was canonical all along.
A Helm chart from the OCI registry for Kubernetes, a hardened multi-architecture container image, a workstation install, and an airgapped path. All of them run against PostgreSQL you manage. The chart and image are public: no registry credentials. Deployment paths →
The authenticated HTTP API ships an OpenAPI 3.1 contract served from the same bytes as the server, so it cannot drift between releases. A dependency-free Python quickstart runs one full authenticated lifecycle. See the API →
Some entries are checked against another source, such as a linked commit being reachable on the declared repository. Others are submitted: notes, evidence references and artifact hashes are recorded as given, with who gave them and when. The record supports review; it does not by itself establish that every reported action occurred as described.
Reports, review verdicts, and receipts are stored as versioned artifacts. Each promoted version is kept; a change creates a new revision. Cite an artifact by key and revision, and that reference continues to resolve to the same content later.
Agent runs are anchored to the identity that executed them and the work they addressed. The ledger records each run, so "who ran what and when?" is a query instead of a reconstruction from session logs.
When you run Hiveram yourself, the record is a PostgreSQL database you hold. Analysts and data engineers can read it with the tools they already use, such as notebooks, BI tools or plain SQL, without a connector from us.
Give analysts a read-only database login of their own. Writes go through the command line, the API or MCP, where identities and the acceptance rule apply.
Each change to a work order is kept with what changed and when, and each version of the request stays readable. Notes, claims and closures carry the identity behind them. That history answers questions such as how often a goal changes after work starts, or how long acceptance takes.
A documented set of stable views for analysis, and a complete export to plain files that needs none of our software, are planned. Until then the schema can change between releases, so pin your queries to a release.
Hiveram keeps the work graph, work orders, authority, bundles and evidence: who asked, what was done, who checked. NeuroRouter decides what slice of that graph enters the live model window. Those layers compose well, but Hiveram does not require them to make sense.
Canonical work graph, bundles, checkpoints and handoff without needing a context proxy.
Live-window shaping, continuity protection, and focused execution without a shared durable work graph.
Hiveram keeps the record. NeuroRouter decides what slice of that record reaches the model right now. Portable bundles are the bridge.
Choose the buy path first, then choose custody. Hiveram can run locally, customer-hosted, or managed. Pricing and deployment mode are separate on purpose.
The workflow page walks through authority modes, the operator loop and portable handoff. The API page carries the OpenAPI contract and a quickstart. Identity says what Hiveram owns and what it does not. Deploy lists the Kubernetes, container, workstation and airgapped paths. Teams that run a security gate have their own page. Plans and the trial key are in Plans and pricing above.