data-centered.com · landscape · 2026-05-02 · v3 (terminology review applied)

Context-Layer Landscape

58 vendors and patterns plotted against the AI-context-layer stack. X = lifecycle stage (schema → feedback). Y = ownership layer (open standard → incumbent). Position tells you when a tool fires and who owns it. The icon next to each name tells you what kind of thing it is — spec, engine, store, graph, memory-system, pattern, platform, or loop. Row D is the analytics-domain moat; the indigo-dotted bubble is the freshness-with-lineage convergence cluster.

your moat — analytics-domain stack

freshness-with-lineage cluster

↓ plugs in

↑ threatens

prepare

apply

learn

01schema / spec

02index / freshness

03substrate / corpus

04memory / recall

05compose

06deliver

07execute

08observe

09feedback / improve

A protocols & conventions open standards · the "buy" tier

DESIGN.md AGENTS.md SKILL.md RDF / PROV-O

MCP

OpenTelemetry

eval-as-spec

B OSS impl self-hostable · low data-risk

ktx

CocoIndex dlt Transformations llm-wikid

Cabinet TrustGraph

ClawVault claude-mem Mem0 Letta Zep

Claude Code Tasks OpenClaw Agent Relay Ramp Inspect Ralph Pattern Taskmaster

E2B / Modal (ref)

Langfuse (ref)

auto-harness Hermes nudges SkillsBench

C vendor SaaS hosted · holds your corpus

Mesa FS Dossium

Glean Hamster Actian VectorAI

Agno

Worklayer Notion MCP Pattern ClawHub

Modal (ref)

Arize Langsmith (ref)

D analytics-domain★ producer-time · governed metrics · the wedge

meta-context schema

dbt-agent canonical models

napkin / beliefs / session-log

Symphony

dbt-mcp

E domain analogs same architecture · other domains · proof-of-pattern

PlayerZero Maximor Regie Fintool

Cognition / Devin (eng correction-capture)

Fintool (mcp)

PlayerZero (telemetry)

F incumbents data gravity · span all columns

spans every column ↑ Salesforce Agentforce Snowflake Cortex Databricks AgentBricks Atlassian Rovo

buy evaluate monitor avoid ★ new this batch (apr 30) faded · ref-only, no canonical entry click any chip → its homepage / canonical doc "gap" cells · row-D whitespace = white space

Icon vocabulary — what kind of thing is each?

spec / protocoldeclares a contract that others implement (DESIGN.md, MCP, OTel, AGENTS.md, meta-context schema, dbt-mcp)

enginedeclarative dataflow / orchestration / runtime (CocoIndex, OpenClaw, Symphony, Agno, Claude Code Tasks, Worklayer, Arize)

store / corpuswhere bits persist — files, blobs, canonical models (Cabinet, Mesa FS, dbt-agent canonical, ClawHub, Fintool corpus)

graphrelational substrate — typed nodes + edges + provenance (TrustGraph, PlayerZero, Maximor, Regie, Dossium)

memory-systemwhat gets remembered, consolidated, forgotten, recalled — retrieval policy, not persistence (ClawVault, Glean, Hamster, claude-mem, Mem0, Letta, Zep, Cognition correction-capture, napkin/beliefs/session-log)

patternpublished architectural recipe — copy not buy (Ralph, Ramp Inspect, Notion MCP Pattern)

platform / incumbentdata-gravity giant spanning all stages (Salesforce, Snowflake, Databricks, Atlassian)

loop / feedbackeval-driven mutation, regression-gated improvement (auto-harness, Hermes nudges, SkillsBench)

v4 changelog — 2026-06-11 substrate-row refresh

Added ktx (row B, schema/spec) — a code-level read of github.com/kaelio/ktx confirmed Apache-2.0 open source (the catalog had it filed as vendor SaaS from docs alone). Schema-validated YAML semantic sources + Markdown wiki, reconciled against the live stack with query-history staleness evidence.
Added dlt Transformations (row B, index/freshness) — dltHub's one-runtime metadata graph from ingestion through transformation; business context riding on it is still prose markdown.
Subtitle count refreshed (31 → 58 plotted chips; the catalog behind the grid is at 37 entries).
Deep-read evidence: thoughts/shared/research/2026-06-11-ktx-dlt-vs-meta-context.md

v3 changelog — gpt-5.4 terminology review applied

Sent to GPT-5.4 (xhigh effort) for architectural review of the taxonomy. Verdict: memory was being smeared across store, graph, and learn — a real competitive surface was being underweighted. Applied:

Added 04 memory / recall as a first-class column between substrate and compose. Reading order: substrate → memory → compose (you persist artifacts first; memory systems compile/retrieve from substrate; composition assembles task context from recalled material).
Added the memory-system archetype, dropped app (row C "vendor SaaS" already encodes the productized signal — redundant icon).
Renamed column 03 store / substrate → substrate / corpus; column 09 learn → feedback / improve.
Renamed row A protocols & specs → protocols & conventions; Y-axis renamed from "specialization" to "ownership layer".
Reclassified misplaced entries: Claude Code Tasks (A → B), Taskmaster (C → B), Cognition / Devin (schema → memory), Fintool (compose → substrate).
Row D gains napkin / beliefs / session-log in the new memory column — the librarian's actual architecture, now visible as part of the moat.
Memory column populated with referenced-but-not-yet-cataloged chips: claude-mem, Mem0, Letta, Zep (faded — pending landscape yaml entries).

// review at thoughts/shared/reviews/context-layer-landscape-terminology/response-gpt-5.4.md

// tokens linted via DESIGN.md · animation rules via emil-design-eng · companion views: cards → · relationships →