Loxtep is an Enterprise Context Layer — here's what that actually means

Why a new category matters now

Every team wiring AI into their operations hits the same wall: the model is capable, but the context it needs is scattered, inconsistent, and ungoverned. You can connect MCP tools, load docs, and wire up RAG — and the AI still invents relationships, contradicts itself across sessions, and forgets what it learned yesterday. The issue isn't model capability. It's the absence of a coherent, governed context substrate that agents can query instead of reinventing.

The Enterprise Context Layer (ECL) is the name for the system that solves this. Prukalpa at Atlan articulated the framework in What an Enterprise Context Layer Actually Is: a context substrate in three parts, activated through five capabilities, governed so it compounds rather than decays. When we evaluated that framework against what Loxtep already ships, we found we'd built the substrate and two of the five capabilities — and had clear primitives for the other three. So we're adopting the category and closing the loops.

The Context Substrate: three parts Loxtep already ships

The substrate is the machine-usable substance of context — not ephemeral prompt fragments, but durable, governed, queryable knowledge. The framework describes three integrated parts. Loxtep ships all three.

1. AI-Ready Data + Knowledge Graph

What the framework says: Governed, structured data assets and their interconnections — the factual foundation AI reads from and reasons over.

What Loxtep ships: Governed data products (source and consumer) on the rstreams streaming backbone. An entity context graph where get_entity_context returns everything about a business entity — across systems, already correlated. Catalog discovery (run_discovery, search_catalog) so data is findable. Lineage and impact analysis so you know where data came from and what breaks if it changes.

2. Semantics + Ontology

What the framework says: The meaning layer — definitions, relationships, canonical keys, namespaces, and vocabulary that let AI resolve ambiguity and speak the organization's language.

What Loxtep ships: A semantic layer with searchable business definitions and metrics. A lexicon (per-data-product glossary). A thesaurus mapping canonical keys to system-specific aliases (resolve_canonical_key). Ontology concepts and relationships. Namespace mappings for multi-system vocabularies. A connector vocabulary inference pipeline that detects synonyms across newly connected sources.

3. Skills (Procedures + Norms)

What the framework says: Encoded procedural knowledge (“how work gets done”) and the access norms that constrain agent behavior.

What Loxtep ships: Two complementary skill types. Organizational Skills — reusable, versionable procedures in the process graph (PKO namespace). They encode how-to: steps, decisions, triggers, dependencies. A skill does for procedural knowledge what a function did for logic — and it compounds, because many agents reuse the same procedure rather than re-deriving it. Agent-Scope Skills — scoped access bundles that declare what an agent may touch, enforced fail-closed. Together: what's possible and what's permitted.

Decision traces record what agents actually do — which procedures they follow, which decisions they make, what the outcomes are. That's the raw signal the learning loop (below) will promote into durable context.

The Five Capabilities: all shipped

The framework defines five capabilities that activate the substrate — turn static context into a living, improving system. All five are now shipped and live.

✅ Activation & Retrieval — shipped

One governed foundation, consumed through many delivery formats. Loxtep doesn't force consumers onto a single interface. The same context is available via MCP (the hosted tool surface AI agents call), REST/API, SQL/Analytics (DuckDB), Webhook/Streaming (rstreams events), Typed SDK (@loxtep/sdk + CLI), and Graph/Entity Context. You integrate where your stack already lives; the underlying knowledge stays canonical.

✅ Governance & Observability — shipped

The controls that keep context trustworthy: RBAC with fail-closed scope enforcement, quality rules with testable assertions, lineage and impact analysis, PII tagging and governance classification, and catalog-wide governance flags. Context without governance decays into the same mess that made data lakes unusable. Loxtep governs by default so the knowledge layer stays reliable as it scales.

✅ Context Development Lifecycle (CDLC) — shipped

A managed lifecycle for context artifacts: draft → in_review → approved → deployed → retired, with dependency tracking and change propagation policies. When a core definition changes, the system knows which downstream artifacts are affected and applies the configured action — auto-propagate, queue for review, or freeze until certified. Built on existing primitives (schema versioning, workspace versioning, lineage impact), unified under a cross-artifact lifecycle with explicit propagation rules via the loxtep_cdlc MCP facade.

✅ Context Mining — shipped

AI-assisted reverse-engineering of business operations from connected systems and runtime signals. Mining detects semantic conflicts (same canonical key, disagreeing definitions), discovers undocumented procedures from event sequences and decision traces, and surfaces them as candidates for human approval. The procedure inference engine (procedure-inference.ts) and the connector vocabulary inference pipeline do the heavy lifting; the unified candidate lifecycle and loxtep_context_mining MCP surface make it operational. Strictly human-in-the-loop: nothing auto-commits.

✅ Compounding Learning Loops — shipped

The mechanism by which episodic experience (decision traces) is promoted into durable semantic or procedural memory. When a pattern recurs — same procedure, same decision point, same outcome — it becomes a promotion candidate. Once a human certifies it, it becomes a governed Organizational Skill that every future agent inherits. The promotion pipeline routes certified patterns through the CDLC into durable context, observable via the Compounding Metric (certified_procedures_over_time).

The compounding story: your learning loop is your IP

Satya Nadella recently articulated a thesis that every company must build two forms of capital: human capital (knowledge, judgment, relationships, pattern recognition) and token capital (the firm's owned AI capability that improves over time). The real opportunity isn't picking the best model — it's building a learning loop where human and token capital compound. A company should be able to swap out a generalist model without losing the “company veteran” expertise built into their learning system.

Loxtep's Enterprise Context Layer is the infrastructure that builds token capital:

• Governed data products are the institutional memory — structured, versioned, governed.
• The CDLC is how that memory is versioned and propagated across dependents.
• The compounding learning loop (decision traces → promotion → durable procedures) is the mechanism by which every interaction makes the layer smarter.
• Model-independence is guaranteed because context lives in your instance, not in any one model's weights.

The context layer is the “hill climbing machine” — the asset that compounds regardless of which generalist model you wire up. Unlike a fine-tuned model (locked to one provider, expensive to maintain, opaque), an Enterprise Context Layer is governed, portable, and owned by the organization. That's the new IP of the firm.

Why the category matters (and what's coming)

Naming the category matters because it gives teams a mental model for what they're building toward. It's not “a data catalog with AI features” or “an MCP server that does some stuff.” It's a distinct layer in the AI architecture — the context substrate that sits between your systems and your agents, governing what they know and how they learn.

Loxtep ships the full Enterprise Context Layer today: governed data products, streaming, semantic layer, ontology, process graph, entity context, scoped skills, multi-dialect activation, the CDLC with change propagation, context mining with semantic conflict detection, and compounding learning loops with memory promotion. All five capabilities are live.

With all five capabilities shipped, the Enterprise Context Layer is the compounding asset that makes every agent, every workflow, and every decision in your organization demonstrably better than the one before.