1. Восторг и массовый immediate adoption

Много комментариев — “great idea”, “I built one”, “will try this”. Это сигнал, что боль ясна: обычный RAG заново открывает знания каждый раз, а люди хотят накопительную структуру.

Представители:

• #33 @MagicUncleDave — “core Zeitgeist right now”;
• #72 @SeeknnDestroy — агенты и люди должны держать memory up to date;
• #180 @justlovemaki — RAG как “rediscover everything every time” ощущается wasteful;
• #211 @pssah4 — wiki как “pre-compiled intermediate layer”.

2. Взрыв реализаций и self-promo

Комментарии быстро превратились в каталог проектов: CLI, Obsidian vaults, MCP servers, desktop apps, web UI, graph memory, research wiki, code wiki, Confluence-to-MCP.

Примеры:

• #44 @VihariKanukollu — browzy.ai: ingest, compile, query, lint;
• #92 @xoai — sage-wiki: single Go binary;
• #177 @originlabs-app — agent-wiki: slash commands, pure markdown;
• #187 @ethanj — llm-wiki-compiler: incremental rebuilds;
• #228 @iamsashank09 — MCP server for LLM Wiki;
• #523 @qhuang20 — Claude Code plugin with SessionStart hook;
• #613 @kogarashi86 — Confluence tree into Markdown wiki, search index, MCP layer;
• #680 @7xuanlu — desktop app + background daemon.

Сигнал для Trip2G: рынок уже понял “LLM Wiki”. Нужно отличаться не наличием markdown, а operational features: sync, MCP, network/federation, provenance, team workflow.

3. Obsidian / Markdown как human-readable слой

Многие используют Obsidian, но часто говорят, что Obsidian — только viewer/IDE поверх файлов.

Представители:

• #20 @logancautrell — Zed + Obsidian workflow;
• #22 @ppeirce — Dataview/Bases;
• #182 @originlabs-app — “Obsidian is just a viewer”;
• #249 @jurajskuska — “OBSIDIAN COULD BE A STANDARD”.

Trip2G angle: “your files stay Markdown”, но появляется hosted/MCP/network layer для агентов.

4. Index.md не масштабируется

Очень частая тема: один index.md работает на малых базах, но ломается при сотнях/тысячах страниц.

Представители:

• #60 @druce — chunking the wiki, LanceDB;
• #91 @bluewater8008 — classify before extract, token budget for index;
• #98 @mpazik — index helps early but does not scale;
• #125 @tashisleepy — wiki enough only до 500+ docs;
• #232/#233 @singularityjason/@omega-memory — index.md overflows context window;
• #269 @bitsofchris — overfetch, deduplicate, rerank for diversity;
• #304 @shibing624 — TreeSearch / structure-aware retrieval;
• #331 @marktran0710 — BM25 + vector + RRF + LLM evaluator;
• #620 @gulliveruk — retrieval itself becomes a reasoning task;
• #725 @superimpactful — missing routing tables and non-hierarchical taxonomies.

Вывод: Trip2G должен говорить не “у нас есть index.md”, а “у нас есть agent navigation contract”: indexes, search, focused reads, MCP methods, routing, backlinks, graph-like navigation.

5. Provenance, citations, truth maintenance

Самая сильная нерешённая боль: если LLM записал факт в wiki, как доказать источник? Как понять, что факт устарел? Как удалить плохой источник?

Представители:

• #77 @laphilosophia — hardest part is truth maintenance;
• #131 @Paul-Kyle — “git blame on every fact”;
• #132 @Jwcjwc12 — every proposition records source files and content hashes;
• #176 @tomjwxf — synthesis without citations is core unsolved problem;
• #234 @ap0phasi — data provenance nightmare;
• #250 @Marekai — page-level citations and no hallucinated citations;
• #287 @asong56 — hallucinations can become permanently embedded;
• #374 @Shagun0402 — trading ephemeral hallucinations for persistent errors;
• #615 @vincent-pli — how to remove/reverse an ingestion;
• #714 @kytmanov — every claim links back to source pages with content hashes.

Функциональное расширение:

• claims.jsonl или frontmatter для claims;
• source id, source hash, timestamp, paragraph/page span;
• confidence/status: verified, disputed, stale, human-approved;
• reverse ingestion: удалить источник и пересобрать зависимые страницы;
• stale detection по source hash;
• citation accuracy benchmark.

6. Lint, drift detection, contradiction checks

Люди хотят не просто “wiki generated”, а “wiki can be audited”.

Представители:

• #42 @skpalan — periodically audit its own wiki;
• #99 @localwolfpackai — Counter-Arguments & Data Gaps;
• #138 @trox — drift detection plugin;
• #213 @frosk1 — error accumulation & drift;
• #218 @sovahc — cache poisoning risk;
• #381 — multi-category wiki lint;
• #680 — daemon with dedup and contradiction checks;
• #720 — doctor: graph integrity, source hygiene, secret-looking files.

Trip2G feature idea: trip2g doctor / MCP method for:

• broken wikilinks;
• stale pages;
• unverified claims;
• contradictory claims;
• orphan pages;
• missing frontmatter;
• secret-looking files;
• pages without source citations.

7. Typed entities, ontology, graph layer

Многие считают, что markdown-ссылки быстро превращаются в неявную БД, значит нужна схема.

Представители:

• #57 @buremba — entity types, strict schema, event log, Postgres;
• #98 @mpazik — wiki grows a schema on its own;
• #134 @blex2011 — graph database built on ontology;
• #146 @xoai — ontology is the hardest part;
• #261 @anzal1 — confidence-scored claims, temporal tracking, contradiction detection;
• #399 @abbacusgroup — OWL-RL ontology + SPARQL + SQLite FTS5;
• #462 @payneio — wiki is just a view for humans, graphdb behind it;
• #498 @nigelglenday — flat markdown backlinks not enough, typed entities/relationships;
• #540 @skyllwt — typed entities and typed edges;
• #590 @sly-codechum — entity linking, conflict handling, supersession, temporal validity.

Trip2G angle: Markdown как portable artifact, но можно добавить schema layer:

• type: concept/person/org/source/task/claim/decision;
• relations: supports, contradicts, supersedes, depends_on, derived_from;
• MCP methods to search by type/relation;
• graph view / routing table as generated notes.

8. Human-in-the-loop and anti-slop

Сильная критика: если LLM сам пишет wiki без контроля, получается “slop layer”; автосводка не равна мышлению.

Представители:

• #217 @robertandrews — generation effect: active processing matters;
• #219 @Runecreed — slop machine;
• #235 @Yuncun — AI slop layer to maintain;
• #410 @asakin — first N days human reviews writes;
• #412 @jurajskuska — humans are the answer;
• #414 @IlyaGorsky — nothing writes without human confirmation;
• #475/#480 @gnusupport — LLM should refresh, not curate;
• #657 @earaizapowerera — LLM can propose, human authorizes;
• #703 @paulshomo — genuine wiki needs splitting, merging, diffing, versioning.

Trip2G positioning: “anti-slop wiki” — agents propose changes, but durable artifacts have logs, citations, review, diffs, rollback.

9. Multi-agent / team memory

Комментарии видят LLM Wiki как shared memory for agents and teams.

Представители:

• #13 @geetansharora — how to share knowledge base with team;
• #46 @Arrmlet — multiple LLM agents maintaining wiki in parallel;
• #174 @GeminiLight — multi-agent is the real unlock;
• #202 @jurajskuska — specialized agents, sandboxing;
• #215 @pnakamura — 7 specialized agents not getting smarter without shared memory;
• #253 @junbjnnn — repo wiki for PRDs, meeting notes, ADRs, runbooks;
• #364 @vitalii-ivanov-rakuten — team wiki as git submodule, Claude PRs upstream;
• #426 @redmizt — identity, access control, contamination prevention, concurrency;
• #494 @devsarangi2 — same corpus has different value for PM vs engineer;
• #679 @gowtham0992 — works with Kiro, Claude Code, Cursor, Codex;
• #694 @AEVYRA — multi-agent drift across Claude/Codex/Gemini;
• #727 @zhurudong — same file works across Claude Code / Codex / OpenCode.

Trip2G angle: shared MCP-accessible knowledge base with roles, access, logs, agent-neutral contract, review queue.

10. Multimodal ingest and event-driven capture

Люди хотят не только markdown/PDF, но всё, что становится знанием.

Представители:

• #102 @uggrock — PDFs, emails, screenshots, voice memo transcripts;
• #105 @Okohedeki — TikTok, tweets, YouTube;
• #181 @lucasastorian — PDFs, PowerPoints, Word, Excel;
• #195 @liamsysmind — text, files, audio recordings, local voice memo transcription;
• #224 @xoai — PDFs, Word, spreadsheets, PowerPoints, EPUBs, emails, images;
• #393 — 50+ formats;
• #724 @VeniVeci — VLM Wiki for photos/videos.

Trip2G opportunity: adapters/importers as landing pages:

• Obsidian-to-MCP;
• Notion/Drive-to-MCP;
• Slack/Discord archives;
• Telegram channels;
• PDFs and docs;
• YouTube/RSS/sitemap;
• screenshots/voice memos.

Core functional extensions

1. Claim-level provenance

   • Each claim knows source, source hash, quote/page span, timestamp, model run, confidence.

2. Citation-aware PDF/document ingest

   • Page-level citations, paragraph IDs, bounding boxes, original quote snippets.

3. Staleness and reverse ingestion

   • If source changes or is removed, mark derived notes/claims as stale and selectively rebuild.

4. Contradiction and counter-argument layer

   • Automatic sections: disputed claims, counter-arguments, data gaps, unverified claims.

5. Multi-model verification

   • Verifier, devil’s advocate, synthesizer; consensus/voting for important facts.

6. Hybrid retrieval over wiki

   • Markdown as artifact, FTS/BM25/vector/graph as search layer.

7. Progressive index / routing tables

   • Global index → domain index → page TLDR → full page; token-budgeted indexes.

8. Typed ontology and entity resolution

   • Entity pages for people/orgs/papers/concepts/tasks/decisions/sources; typed relations.

9. Human review queue

   • Draft updates, staged candidate buffer, approval before canonical write, rejection feedback.

10. Git-native audit and rollback

• Every change is commit-like: reason, affected claims, source hashes, diffs.

11. Multi-agent write coordination

• Task claiming, locks, roles, permissions, serial write daemon, branch/PR workflows.

12. Multimodal/event-driven ingest

• Watch folders, mobile capture, share sheets, voice memo ingestion, screenshots, browser history.

13. Benchmarks/evals

• Fixed corpora, recall@k, citation accuracy, contradiction detection, freshness, token cost.

14. Security scanner

• Detect secrets/PII/private docs before publishing or agent exposure.

Рекомендуемое позиционирование

Trip2G is the hosted MCP knowledge-base layer for LLM Wikis: your Markdown stays human-readable, while agents get search, focused reads, sync, provenance, logs, and safe update workflows.

Или короче:

LLM Wiki is the artifact. Trip2G is the MCP/network layer that lets agents read, update, cite, sync, and federate it.

Что стоит добавить в roadmap / landing

Now:

• Docs/landing page: “Run Karpathy-style LLM Wiki as an MCP-accessible hosted Markdown vault”.
• Example vault: raw sources + wiki pages + _index.md + AGENTS.md + 99 - Журнал действий.md.
• MCP quickstart: search → note_html → update via sync/API.
• Navigation docs: _header.md, _footer.md, wikilinks, index-first traversal.

Next:

• SOURCES.md / sources/ convention.
• claims.jsonl or frontmatter schema for claim provenance.
• trip2g doctor: broken links, stale pages, missing citations, secrets scanner.
• Focused reads and match IDs in MCP docs.
• Review queue pattern: draft note → human-approved canonical note.

Later:

• typed entity graph over markdown;
• reverse ingestion;
• source-hash stale rebuild;
• role-aware compiled views;
• multi-agent locks/permissions;
• federation across Trip2G bases.

Короткий комментарий

This pattern strongly matches what we are building with Trip2G: a hosted, MCP-accessible layer for Karpathy-style LLM Wikis over plain Markdown/Obsidian vaults.

The wiki stays the durable artifact, but agents can search it, read focused pages, update it through sync/API, keep logs/provenance beside the content, and later federate multiple wikis into a knowledge mesh.

I think the next hard parts after “LLM writes markdown” are exactly what many comments here mention: provenance for claims, stale-source detection, reviewable updates, multi-agent coordination, and navigation beyond one index.md. We are exploring those as conventions like AGENTS.md, SOURCES.md, logs, wikilinks, MCP search/read methods, and sync workflows.

If anyone is building LLM Wiki tooling and wants an MCP/hosted Markdown layer to interoperate with, I’d love to compare notes: https://trip2g.com/

Более прямой комментарий с позиционированием

Karpathy’s framing “the wiki is the artifact, not the chat” is exactly the direction I think agent memory is moving.

One thing we are experimenting with in Trip2G is making that artifact networked and agent-readable: plain Markdown/Obsidian vaults published as a site + exposed through MCP, so Claude/Codex/Cursor-style agents can search, read, cite, and update the wiki instead of keeping knowledge trapped inside a chat/session.

A lot of comments here point to the same next layer: provenance, stale facts, review queues, multi-agent writes, and better navigation than a single index.md. My current hypothesis is that LLM Wiki needs an operational layer: sync, logs, source conventions, MCP tools, and eventually federation between knowledge bases.

Would be happy to connect with people building in this direction: https://trip2g.com/

Самый короткий вариант, чтобы не выглядело как реклама

One extension I’d love to see standardized: an MCP contract for LLM Wikis.

If the wiki is the durable artifact, agents need a common way to search it, read focused pages, inspect sources/provenance, and propose updates safely. We’re exploring this in Trip2G with plain Markdown/Obsidian vaults exposed as hosted MCP-readable knowledge bases: https://trip2g.com/

Рекомендация: использовать самый короткий вариант. В треде уже много self-promo, поэтому лучше зайти через идею стандарта MCP contract, а Trip2G дать как один пример.