Noma Cloud Research -- Agent-Human Knowledge Platforms

The market has moved beyond adding a chat box to a wiki. Confluence, Notion, Glean, Guru, and Slite now compete on permission-aware search, cited answers, connected sources, governed agents, and knowledge freshness. The remaining opening for Noma is sharper: a block-native knowledge system where retrieval returns exact source-backed blocks, every answer exposes versioned citations, and every agent write becomes a proofed patch that a human can inspect and approve.

Research scope

This memo reviews public product material and user feedback available on 2026-07-14 for:

  • Atlassian Confluence and Rovo
  • Notion, Notion AI, Custom Agents, and the developer platform
  • AI-native knowledge products including Glean, Guru, and Slite
  • recent Confluence and Notion user reviews and feature-request discussions

The target audience is technical, research, product, and documentation teams. The time horizon is the next 12 to 24 months. Product pages establish shipped capabilities; review aggregations and community threads provide directional pain signals. Anecdotes are treated as evidence of a problem shape, not as market-size estimates.

Executive read

The competitive baseline is therefore higher than "RAG inside the wiki." Users expect answers to respect permissions, cite sources, and reflect current content. Admins expect agent scopes, activity trails, cost controls, and the ability to stop unsafe automation. Authors still report slow large workspaces, hard navigation, confusing permissions, stale duplicates, and setup complexity.

Noma should not copy the breadth of Confluence or Notion. It should use the assets those products do not have as a coherent contract: readable source, stable block IDs, typed knowledge, block hashes, validation, deterministic LLM context, and proof-before-apply patch operations.

Market baseline

ProductCurrent center of gravityStrong market signalOpening for Noma
Confluence + RovoJira-connected wiki, enterprise search, chat, agents, studio, MCP partner skillsDeep work graph and action across Atlassian toolsSimpler source-first workspace with transparent block-level lineage and lower governance complexity
NotionFlexible docs/databases plus enterprise search, agents, external-agent orchestration, and developer platformExcellent authoring flexibility and fast agent platform expansionDurable portable source, safer writes, clearer knowledge contracts, and less workspace sprawl
GleanPermission-aware enterprise search and knowledge graph across many systemsBest-of-breed cross-source retrieval contextAuthoritative creation and maintenance layer, not only a search overlay
GuruVerified knowledge, cited answers, knowledge agents, MCP deliveryTrust, verification, and auditabilityTyped evidence and versioned block proofs that remain portable outside the service
SliteSelf-maintaining AI knowledge base with Ask/Agent, verification, and block editsTight loop from answer to document maintenanceOpen source-backed format plus rigorous agent change review

What users keep asking for

Confluence pain clusters

G2's 2026 review aggregation identifies navigation/search friction and slow performance as repeated themes, each with 83 mentions; difficult page management appears in 70, learning curve in 64, and overwhelming complexity in

  1. Recent individual reviews add stale or duplicate search results, confusing

permissions, notification fatigue, plugin dependence, and rising cost. These counts describe G2's review corpus, not the whole market, but the clustering is consistent enough to guide product priorities.

ProblemUser goalWhat breaksNoma response
Find the current answerRetrieve one trustworthy page or block quicklySearch returns stale duplicates and deep page trees hide contextHybrid retrieval ranked by permission, verification, freshness, type, and graph relationships
Keep the space understandableAdd knowledge without creating a mazeNested pages, macros, and inconsistent naming produce sprawlOpinionated spaces, typed templates, automatic link suggestions, orphan/duplicate detection
Share safelyUnderstand who can see and change knowledgeLayered permissions are easy to misconfigureExplainable effective-access view plus small role vocabulary and agent-specific scopes
Work at scaleOpen and edit large spaces predictablyHeavy pages and large spaces become slowBlock-indexed reads, progressive loading, bounded views, performance budgets
Maintain knowledgeKnow what is stale, unused, or duplicatedGovernance depends on discipline or add-onsVerification owners, review-by dates, supersession, and automatic health queues

Notion's most requested feature themes

Offline access was the clearest long-running request. Notion itself described the August 2025 offline release as the feature people had been waiting for, but said only essential blocks worked initially. 2026 reviews still describe limited database synchronization and unreliable access to some offline content. That distinction matters: the headline feature shipped, while the reliability and coverage request remains open.

Granular permissions are another persistent theme. Notion added a "can create pages" database permission in March 2026, allowing users to create records without seeing other restricted entries. Community requests still ask for clearer row, column, property, and shared-view boundaries. Recurring tasks, performance with large databases, faster mobile capture, simpler setup, better search in grown workspaces, richer API throughput, and tighter Calendar/Mail integration recur across recent reviews and request threads.

Request themeStatus by July 2026Frequency signalProduct lesson
Offline/local reliabilityShipped in 2025, still partial for some blocks and database depthHistorically the loudest request; still present in 2026 reviewsOffline is a trust property, not a checkbox
Granular permissionsImproving, including 2026 create-only database accessRepeated in team-focused request threads and reviewsUsers need comprehensible boundaries more than dozens of opaque roles
Performance at scaleOngoing improvement, still reported for large pages/databasesRepeated in G2 reviewsFlexibility without performance budgets becomes product debt
Recurring work and schedulingImproved, but advanced recurrence remains requestedRepeated in 2025 community threadsKnowledge maintenance needs native schedules and triggers
Simpler setup and navigationTemplates and AI help, but flexibility still overwhelmsG2: learning curve 1,907 mentions; limited features 1,032; unintuitive setup 944Progressive disclosure and opinionated defaults beat a blank canvas
API and cross-app continuityDeveloper platform expanding rapidlyReviews still request higher limits and tighter Mail/Calendar/database flowAgents and integrations must be first-class, observable clients

Ranked product problems

ProblemSeverityFrequency signalConfidenceRecommended move
Untrusted or stale answersCriticalHigh across wiki reviews and AI-native positioningHighBuild verification, freshness, supersession, citations, and answer refusal before broad autonomy
Search returns documents, not usable contextHighHighHighRetrieve typed blocks with surrounding section, graph links, and exact source/version
Agents can read but safe write-back is weakCriticalGrowing platform-level signalHighMake plan -> proof -> independent approval -> apply the universal agent write path
Knowledge sprawl and duplicate truthHighHigh in both Confluence and Notion feedbackHighAdd canonical pages, duplicate suggestions, orphans, missing concepts, and merge proposals
Permission behavior is hard to reason aboutCriticalHigh for team useHighProvide effective-access explanations and use the same filter for search, RAG, exports, and agents
Large workspaces become slow and noisyHighHighMedium-highSet page/block/query budgets and measure latency before adding realtime editing
Automation lacks governanceHighRising as agent directories growHighAdd scoped agent identities, run logs, budgets, model policy, triggers, and kill switches
Offline and portable continuity remains fragileMedium-highPersistent Notion signalMedium-highPreserve export/Git strengths now; research local-first sync after server correctness

Proposal portfolio

P0 -- trust and retrieval foundation

  1. Block-native hybrid RAG. Combine SQLite FTS, embeddings, typed block

metadata, page/space graph edges, freshness, verification, and access checks. Index canonical source blocks rather than anonymous copied chunks. Retrieval results must include document ID, block ID, source span, version hash, visibility decision, and neighboring context.

  1. Ask Noma with versioned citations. Add a workspace question surface and

API that returns a short answer, exact block citations, confidence, and a visible "insufficient evidence" state. Opening a citation should highlight the source block and the version used to answer.

  1. Knowledge trust graph. Add owner, verified-by, verified-at, review-by,

supersedes, canonical-for, and source-of fields around existing Noma blocks. Verified/current sources rank above unverified/stale ones. Conflicts should be shown, not silently averaged by the model.

  1. RAG evaluation harness. Let a space keep golden questions with required

sources and forbidden sources. Measure retrieval recall, citation coverage, permission leakage, stale-source use, answer abstention, latency, and cost before changing models or index settings.

  1. Knowledge health queue. Generate actionable queues for stale pages,

missing owners, orphan pages, broken links, duplicate candidates, contradictory claims, unanswered high-frequency queries, and documents an agent used despite low trust.

P0 -- LLM Wiki and safe collaboration

  1. LLM Wiki mode. Extend existing page wikilinks/backlinks with suggested

links, missing-concept pages, canonical concept pages, typed relationships, and duplicate merge proposals. Agents may draft or reorganize pages, but every generated block carries provenance and every write remains a patch proposal.

  1. Agent change inbox. Turn the existing proofed patch-proposal API into a

shared review queue. Show agent plan, requested capabilities, sources used, diff, validations, affected IDs, reviewer, and apply status. Keep self- approval blocked for material changes.

  1. Human/agent presence model. Treat an agent as a scoped collaborator with

its own identity, page/space access, model, budget, and action capabilities. Never inherit the triggering human's full access implicitly.

P1 -- connected and self-maintaining knowledge

  1. Connector framework with source lineage. Start with GitHub, Slack, Google

Drive, Jira/Linear, and filesystem imports. Preserve upstream permission, modified time, deletion/tombstone, and source URL. Avoid a shadow vector store that loses the source system's access semantics.

  1. Agent directory and recipes. Ship narrow agents for stale-doc review,

meeting-to-decision conversion, issue-to-runbook updates, research refresh, onboarding answers, and release-note maintenance. Support manual, scheduled, event, and webhook triggers with per-run logs and budgets.

  1. Semantic collections and views. Let users query typed blocks across

pages -- all open decisions, claims missing evidence, risks by owner, agent changes awaiting review -- without moving the source of truth into an opaque database.

  1. External agent gateway. Expose permission-aware search, cited answer,

IDs, proof, proposal, and approval through MCP/API/webhooks. External agents should receive the same scoped context and safety gates as native agents.

  1. Search and answer analytics. Record no-result queries, reopened search

sessions, citation opens, answer rejection, and successful task completion. Use aggregates to prioritize missing knowledge; never expose hidden query content across permission boundaries.

P2 -- continuity and enterprise depth

  1. Git and backup bridge. Provide scheduled .noma export/import with

deterministic paths, version hashes, conflict reports, and optional pull request review. Portability is a product feature and a hedge against knowledge-platform lock-in.

  1. Offline-capable app. Research a local-first/PWA cache only after save,

history, and merge semantics are explicit. The acceptance bar is full draft recovery, deterministic conflict presentation, and predictable database/ asset coverage.

  1. Realtime co-editing. Add presence and multiplayer editing only when it

can preserve block IDs, version history, proof preconditions, and portable .noma source. Async proofed collaboration remains the safer default for agent-authored changes.

  1. Enterprise governance. SSO/SCIM, retention, legal hold, data residency,

audit export, connector allowlists, model allowlists, zero-retention model options, and per-agent spend controls belong after the core trust loop is proven.

Product rules

Source map

Strongest capability evidence came from first-party Atlassian, Notion, Glean, Guru, and Slite product/help material. G2 supplied broader review clusters and recent individual examples. Reddit and community threads helped identify feature-request language, especially around offline mode, permissions, and recurring work, but those anecdotes are not used as quantitative market shares. No internal customer interviews, support tickets, or Noma usage analytics were available; the next research step should validate the top five problems with five to eight target teams before committing to connector or realtime work.

Source notes

Atlassian overview of Rovo search, chat, agents, studio, and connected knowledge inside Confluence.

Atlassian's 2026 direction for Confluence partner agents, MCP skills, and source context carried through the Teamwork Graph.

G2 review aggregation for Confluence, including repeated navigation/search, performance, page-management, learning-curve, and complexity themes.

Notion Enterprise Search documentation covering workspace, connected-app, web, and cited answer retrieval.

Notion release stream covering Plan Mode, Custom Agent admin controls, the agent directory, developer platform, and External Agents orchestration.

Notion documentation for agent tools/access, model choice, permissions, activity, version history, triggers, and maintenance.

Notion's offline-mode release, described by the company as the feature users had been waiting for and initially covering the most essential blocks.

Notion's create-only database permission, allowing new entries without access to other restricted records.

G2 Notion review aggregation and recent review examples covering learning curve, missing features, large-workspace performance, permissions, search, offline limitations, API limits, and cross-app continuity.

2025 community feature-request thread with recurring requests for offline mode, granular database permissions, automation improvements, and recurring tasks.

Guru product overview for verified knowledge, permissions, citations, audit lineage, MCP delivery, and agent actions in connected systems.

Guru documentation for automated verification signals and verification badges in search and AI-generated answers.

Slite's self-maintaining knowledge-base direction, including agent search, block edits, verification, archive, comments, and connected sources.

Slite documentation describing verified documents ranking above outdated ones in search and AI answers.

Glean product overview for permission-aware enterprise search, agents, and cross-company context.

Glean documentation describing its real-time enterprise knowledge graph and permission management foundation.

Interview five to eight technical, research, or product teams that currently use Confluence, Notion, or both. Test the ranked problems, ask for recent failed search examples, and measure willingness to adopt block-cited answers plus proofed agent write-back before expanding the roadmap.