Skip to content

Team-Scale Knowledge Tooling — Competitive Landscape & Lessons for AsDecided

Status: research document, not a RAC artifact. Per ADR-010 ("Documents Are Not Artifacts") and ADR-024 ("RAC Is Not a Content Store"), this is reference / working material, not part of the validated decisions/ corpus. Its actionable conclusions are captured as corpus artifacts — see "What this produced" below.

Scope: how software teams of 20+ engineers (not indie hackers) adopt, govern, and sustain knowledge tools, and what AsDecided (RAC) should learn. Method: multi-angle web research across four adjacent product categories plus a cross-cutting failure-mode pass, with adversarial verification and evidence-vs-marketing tagging. Confidence and primary sources cited inline. Date: 2026-06.

Executive summary

Across every category the finding is the same: capture is solved and identical everywhere (Markdown-in-git, immutable records, status + supersession). What decides survival at 20+ is (a) fighting staleness/trust-collapse, (b) discovery, and (c) riding the existing workflow. AsDecided's no-database / files-in-git / PR-review / MCP design is structurally aligned with nearly every sustaining mechanism the evidence supports — its single biggest exposure is the freshness-signal gap, which is both the best-documented cause of abandonment and the gap no agent-context competitor has solved.

One-sentence positioning that the research supports:

AsDecided is the shared, git-native, deterministic source of truth for product decisions — the knowledge class no agent-context tool governs and no team can afford from enterprise requirements management — winning on the freshness/drift problem (now empirically the #1 adoption determinant) and the no-embeddings bet that a major incumbent (Sourcegraph) publicly reversed into.

Category findings

1. AI agent context / memory for teams

Cursor, Continue, Sourcegraph Cody & Amp, Augment, Glean, Dust, Onyx.

  • Files-in-git + PR review is now an industry pattern, not a compromise. Cursor (.cursor/rules/*.mdc) and Amp (AGENTS.md) are git-committed Markdown reviewed through the normal PR flow — AsDecided's exact substrate. [high] (Cursor rules, Amp)
  • Sourcegraph publicly abandoned embeddings for enterprise context (third-party code transmission, vector-DB maintenance, poor scaling) and replaced them with deterministic search — strong external validation of AsDecided's no-embeddings stance (ADR-066). [high] (Sourcegraph)
  • "Shared team memory" is oversold. Cursor Memories and Augment's base index are per-user; real org memory (Augment Cosmos) is preview-only — a genuinely shared, governed team source is unmet demand. [high]
  • Hosted RAG has under-disclosed staleness windows everywhere (Glean up to a ~30-day full-crawl gap; Dust 1–2 days; Onyx 30-min default). "Real-time" is marketing; ACLs gate retrieval, not LLM-synthesis oversharing. [high/medium]
  • The dominant adoption-killer is context being silently ignored by the model — "Cursor is a predictive engine, not a policy enforcer." Shared context buys consistency-by-default, never compliance-by-guarantee — exactly the boundary ADR-067 already encodes (context-supply + post-edit enforcement). [high]
  • None of them ship freshness, expiry, or ownership metadata on the rules themselves — the gap AsDecided can own. [high]

2. Decision & spec tooling

ADR tools (adr-tools, MADR, Log4brains, Backstage ADR plugin); spec-driven dev (GitHub Spec Kit, AWS Kiro, Tessl); enterprise RM (Jama, IBM DOORS, Siemens Polarion).

ADRs. Convergent design since Nygard (2011): Markdown-in-repo, immutable, status + supersession. Tools differ on discovery, not capture (Log4brains adds a static searchable site; Backstage couples ADRs to the catalog entity via backstage.io/adr-location). The documented death pattern is the "abandonment curve" (5 ADRs in month one, then nothing) and stale "Accepted" records eroding trust; the antidote is an operating model (ownership + cadence + a "Gardener" role + bounded review windows + CODEOWNERS), not tooling. ThoughtWorks put lightweight ADRs in Adopt and the rule is "store in source control… in sync with the code." [high on mechanics, medium on failure-mode blogs] (Nygard, MADR, Log4brains, Backstage, ThoughtWorks)

Note for AsDecided: Log4brains deliberately avoids fixed numbered filenames to prevent multi-dev merge collisions — AsDecided's minted opaque IDs (RAC-…) already solve this. And decided review/doctor/coverage automate the "Gardener" role that the evidence says is the manual antidote to the abandonment curve.

Spec-driven dev. Spec Kit / Kiro / Tessl invert the source of truth ("intent is the source of truth, code is generated") via a Markdown phase pipeline. Enforcement is AI-mediated and human-gated, not deterministic (Spec Kit has no blocking gates; Kiro's "neuro-symbolic" analysis uses LLM + SMT). Their central, openly unsolved problem is spec↔code drift — there is an "Ask HN: specs go stale" thread and no vendor documents a deterministic anti-drift mechanism; the recurring critique is "SDD is waterfall." Kiro independently uses EARS (which AsDecided validates deterministically). [high] (Spec Kit, Kiro, Kiro deep analysis, Tessl, "waterfall" critique)

Note for AsDecided: AsDecided sidesteps SDD's core unsolved problem — it doesn't generate code from artifacts, it records decisions an agent consults, so there's no spec→code sync to rot. And it does deterministically the structural slice Kiro needs an LLM+SMT solver for.

Enterprise RM (Jama / DOORS / Polarion). Full formal governance — e-signatures, FDA 21 CFR Part 11, baselines, workflow states, and "suspect links" (downstream items flagged the moment an upstream item changes; DOORS: valid/invalid/suspect). The gap is not capability — it's friction: opaque seat pricing (~$400–600/user/yr DOORS; ~$1,788/user/yr Polarion; Jama quote-only), steep admin, performance that degrades at scale (DOORS Next is vendor-acknowledged poor; Polarion is bottlenecked by SVN single-writer commits), brutal export/migration lock-in, and — most telling — license cost literally throttles collaboration ("only a limited number of people" get authoring licenses). [high on friction, medium on third-party pricing] (Jama, DOORS link validity, DOORS Next perf, Polarion)

Note for AsDecided: "suspect links" is the proven enterprise drift mechanism — AsDecided can deliver a deterministic, git-native version (commit diffs + the validated graph) without the per-seat tax, SVN bottleneck, or lock-in. Their fatal flaw (friction/cost/lock-in, not capability) is AsDecided's wedge.

3. Internal developer portals

Backstage, Port, Cortex, OpsLevel.

  • Ownership is a committed field for routing, not enforcement (catalog-info.yaml spec.owner); Backstage's own docs say it is explicitly not for runtime auth. [high] (Backstage descriptor)
  • Scorecards = deterministic checks grouped into maturity levels (OpsLevel bronze/silver/gold; a baseline check flags orphaned/ownerless services) — AsDecided's decided validate/review analog. [high] (OpsLevel)
  • Drift is the real enemy; portals fight it with auto-discovery + orphan/drift detection (Backstage backstage.io/orphan + orphanStrategy). AsDecided can't auto-discover (no crawler/DB) but counters with git-derived staleness (ADR-045). [high] (Backstage life-of-an-entity)
  • TechDocs = Markdown in the same repo as code, updated in the same PR — AsDecided's exact model, proven at Spotify scale (280+ teams, 2,000+ services). [high]
  • Adoption: "usage can be mandated, but adoption must be earned" — top-down mandates correlate with lower developer satisfaction; portals become shelfware when devs revert to Slack. [medium, analyst/vendor]

4. Team knowledge bases / wikis

Notion, Confluence, Stack Overflow for Teams, Slab, Slite.

  • The decay problem is real and vendor-acknowledged. Failure modes: content decay ("the gap between documented truth and operational reality widens silently"), ownership vacuum ("everyone is responsible = nobody"), and misaligned incentives ("nobody gets recognized for updating a Confluence page"). [high on mechanisms]
  • Every tool ships the same anti-staleness primitive: owner + expiry clock + freshness signal that biases retrieval (Notion page verification, Slab/Confluence verified status with auto-un-verify, SO for Teams Content Health, Slite verification). But it's a self-clicked "verify" button — human attestation, not proof. [high] (Notion, SO for Teams)
  • Findability decays independently (recycled but directional stats: ~1.8 hrs/day searching; ~½ of searches end in "not found"); knowledge fragments across Slack/docs and the single source of truth splinters. [medium]
  • All five are bolting LLM Q&A over the corpus; Slite explicitly positions verified docs as "the source of truth for Claude, ChatGPT, Cursor." AI surfaces gaps but doesn't fix decay. [high]

Note for AsDecided: its freshness primitive is structurally stronger — git history is deterministic recency (ADR-045) and PR review is enforced attestation — "verification you can't fake" — plus decided validate/relationships as machine-checkable rot detection no wiki has. The inherited risk: PR review raises contribution cost — the exact friction that kills wikis at 20+ (mitigate with agent-assisted authoring + fast gates).

Cross-cutting evidence (the empirical backbone)

  • Docs measurably drift. A study of 27,772 PRs across 714 repos: only ~0.8% of PRs update the README; ~21.5% of changes that should have updated it didn't; doc updates are ~1% of PR activity and routinely overlooked. [high] (arXiv 2603.00489)
  • Documentation quality has a measured payoff. DORA: high-quality docs make teams "more than twice as likely" to hit performance targets, and amplify every technical practice. DORA 2025 explicitly endorses "connecting AI tools to internal documentation, codebases, and decision logs" — AsDecided's exact use case, named. AI is an "amplifier" that magnifies good docs and exposes bad. [high / medium-high] (DORA docs, DORA 2025)
  • The load-bearing adversarial finding: git + PR review alone does not keep docs fresh. The docs-as-code literature itself concedes "no empirical data validates whether drift actually decreases at scale," and the 0.8% data shows reviewers routinely miss doc drift. Co-locating in git is necessary but proven-insufficient — which is why the deterministic freshness/drift signal matters. [high]
  • Context rot is established. Chroma Research tested 18 frontier models; every one degrades as input length grows, well before window overflow — favoring curated, minimal, on-demand context over corpus dumps. [high] (Chroma)
  • The MCP "context tax." Simon Willison stopped using MCP for coding agents because tool descriptions eat context (GitHub's MCP ≈23k tokens) and CLIs are lower-tax — but he endorses the underlying "context engineering" discipline. A knowledge server justifies itself only if it stays lean. [high] (Willison)
  • Honest gap: there is no published proof that curated decision context lifts coding-agent task success. The strongest defensible claim is that context-rot + context-engineering evidence supports selective delivery. AsDecided's grounding-eval (ADR-066) is the way to be the one who proves it. [high confidence in the gap]

What determines whether a 20-person team adopts vs abandons AsDecided (impact-ordered)

# Lesson Tag Action
1 Lives in the PR/git workflow, not a separate destination FITS (strongest) This is docs-as-code — the #1 sustaining mechanism. Lead with it.
2 Staleness → trust collapse is the top killer; git+PR review is proven-insufficient RISK (no DB; recency from git, ADR-045) Surface git-derived staleness loudly + a deterministic "suspect" drift gate. Highest-leverage build. → future/freshness-and-drift-detection
3 Automate the "Gardener" (the manual health-review that's the documented ADR antidote) POSITIONING decided review/doctor/coverage = machine-checkable health no wiki/portal has. Make it a scheduled CI report.
4 "Verification you can't fake" POSITIONING Enforced PR review + immutable git history vs competitors' self-clicked verify buttons.
5 Effortless consumption — context comes to the agent FITS The read-only MCP server pushes context in-flow; counters the "go to the wiki" abandonment driver.
6 …but keep the surface LEAN (context tax + context rot) POSITIONING / RISK Few tools, small descriptions, retrieve-on-demand (ADR-033). CLI-first (ADR-005) may out-adopt MCP. → future/lean-context-delivery
7 Discovery/proximity is the survival differentiator, not capture FITS Attach decisions to the code/services they affect + the typed graph (ADR-074). → future/decision-to-code-proximity
8 Contribution friction from the PR gate (the friction that kills wikis) CONFLICT (ADR-065) Mitigate with agent-assisted authoring (rac-artifacts skill) + fast gates.
9 AI changes the ROI of curation POSITIONING The timing tailwind. Lead GTM with "your agents improve on day one from artifacts you already have."
10 Only wins if it REPLACES a silo, not adds one POSITIONING Position as the source of truth, not "another store" beside Confluence/Slack.
11 Low ceremony FITS ADR evidence: heavy templates get abandoned. Guard against template bloat.
12 Prove it HONEST Grounding-eval (ADR-066) — no public proof curated context lifts task success yet; be the one with the number.

Competitive landscape & positioning

  • Closest single competitor concept: mcp-adr-analysis-server — an MCP server that serves/analyzes ADRs to agents. AsDecided differentiates by being full artifact families
  • deterministic validation + a validated graph, not ADR analysis alone. (repo)
  • Patterns to borrow: Spec Kit's constitution (governed project principles), Backstage's catalog-coupled ADR discovery (adr-location).
  • The whitespace: every agent-context tool stores coding rules or indexed code; none governs product decisions / requirements / roadmaps. That is AsDecided's uncontested niche.
  • The wedge against enterprise RM: the traceability + governance value of DOORS/Jama/Polarion (suspect links, baselines, review), git-native — no per-seat collaboration tax, zero lock-in (it's just Markdown), no SVN bottleneck, deterministic — for teams priced out of or crushed by enterprise RM.
  • AsDecided's defensible middle sits between lightweight MCP file/notes servers (deterministic but unstructured) and enterprise RAG-over-docs (embeddings, precision-poor for decisions): deterministic + structured artifacts + a validated graph.

Honest caveats & open questions

  1. No published proof curated decision context lifts agent task success. Evidence supports selective delivery (context rot), not "decisions demonstrably make agents better." Grounding-eval (ADR-066) is the answer and a moat.
  2. The freshness gap is real and is AsDecided's softest spot — but also the gap nobody else has solved, so it is a differentiator to win, not just a risk to cover.
  3. Decision rationale cannot be auto-generated, so AsDecided can detect and surface staleness, not make it "structurally impossible" the way generated API docs can.
  4. MCP resources are under-adopted (clients favor tools); agent-context conventions (AGENTS.md / CLAUDE.md / Cursor rules) are unsettled — don't over-commit to one surface.
  5. Contribution friction at 20+ is inherited from the PR gate.
  6. Sourcing limits: many adoption/ROI and pricing figures are vendor or third-party estimates (medium confidence); the durable facts are the mechanism docs, the arXiv drift study, DORA, the Sourcegraph embeddings reversal, and the enterprise "suspect link" semantics.

What this produced (captured as corpus)

The actionable findings were distilled into unscheduled future/ roadmap items (this document is the reference behind them):

  • decisions/roadmaps/future/freshness-and-drift-detection.md — loud git-derived staleness
  • a deterministic git-native "suspect links" drift gate (lesson #2/#3/#4).
  • decisions/roadmaps/future/decision-to-code-proximity.md — declared code-scope references so the governing decision surfaces at the point of work (lesson #7).
  • decisions/roadmaps/future/lean-context-delivery.md — measure/bound the agent-facing footprint and keep a first-class CLI path (lesson #6).

The competitive positioning conclusions belong with the rac-growth-positioning requirement or a dedicated design (not yet captured at the time of writing).