Show HN: Harness session memory without transcript hoarding

LLM Wiki, a new open-source tool released as a plugin for Claude Code, OpenAI Codex, and OpenCode, enables parallel multi-agent research and session memory without hoarding transcripts, using human-owned schema.md topic guides and plain Markdown output. The tool dispatches 5–10 agents across academic, technical, applied, news, and contrarian angles, with a --min-time 2h option for iterative rounds, and compiles sources into articles, reports, and study guides while keeping raw sources immutable.

Research 5–10 parallel agents search academic, technical, applied, news, and contrarian angles. --min-time 2h keeps going in rounds, drilling into gaps each round finds. ██╗ ██╗ ███╗ ███╗ ██╗ ██╗██╗██╗ ██╗██╗ ██║ ██║ ████╗ ████║ ██║ ██║██║██║ ██╔╝██║ ██║ ██║ ██╔████╔██║ ██║ █╗ ██║██║█████╔╝ ██║ ██║ ██║ ██║╚██╔╝██║ ██║███╗██║██║██╔═██╗ ██║ ███████╗███████╗██║ ╚═╝ ██║ ╚███╔███╔╝██║██║ ██╗██║ ╚══════╝╚══════╝╚═╝ ╚═╝ ╚══╝╚══╝ ╚═╝╚═╝ ╚═╝╚═╝ Parallel multi-agent research. Thesis-driven investigation. Source ingestion. Wiki compilation. Concept → Idea → Project incubation. Human-owned topic guides. Session memory. Feedback curation. Topic archiving. Inventory tracking. Dataset manifests. Truth-seeking audits. Token-efficient read-only querying. Artifact generation. Ships as a Claude Code plugin, an OpenAI Codex plugin, an OpenCode instruction file, or a portable AGENTS.md . Obsidian-compatible. Every run compounds. Sources become cross-referenced articles. Articles become reports, slide decks, study guides, playbooks, and implementation plans. Session digests keep the agent oriented without turning private chats into topic evidence. One command spins up a topic wiki, dispatches up to ten agents, ingests what's worth keeping, collects provenance-rich catalogs before tracking them, develops rough Ideas into approved Project briefs, archives old topics without deleting them, tracks durable follow-up state, indexes large datasets without copying them, captures redacted session context and feedback candidates without hoarding transcripts, compiles sources into articles, and hands you a deliverable built on top. Human-owned schema.md topic guides keep local vocabulary and conventions explicit. All plain Markdown you own. 5–10 parallel agents search academic, technical, applied, news, and contrarian angles. --min-time 2h keeps going in rounds, drilling into gaps each round finds. Start from a claim. Agents split across supporting, opposing, mechanistic, meta, and adjacent angles. Output is a verdict — not a summary. Round two fights confirmation bias. URLs, files, PDFs, inbox drops, Git doc repos, MediaWiki dumps, message archives, and Wayback CDX snapshots. Raw sources stay immutable; articles synthesize on top. Register explicitly trusted local research tools without publishing their code, credentials, corpora, or bulk results. Manifest handshakes, path scopes, timeouts, and artifact hashes keep execution reviewable; nothing is imported automatically. Find, dedupe, download bounded public media, and catalog discoverable artifacts, examples, memes, tools, entities, and source candidates. Captures aliases, found-in-context provenance, local asset paths, hashes, scale, media policy, and inventory fit. Track durable things the wiki should remember: items, source candidates, corpora, entities, open questions, watch items, and next actions. Chat views default to compact tables. Capture a rough proposal without pretending it is a Project. Research it, challenge it, compare minimal and ideal versions, then explicitly promote an approved snapshot when you are ready to deliver. Index large, external, mutable, or operational data with manifests, samples, profiles, and query recipes. The wiki becomes the interface; the data stays where it belongs. Every new wiki gets a human-owned schema.md topic guide for local vocabulary, relationships, source boundaries, and adoption notes. Older wikis can adopt one safely with llm-wiki schema adopt . Move whole topic wikis to topics/.archive/ . Preserved knowledge stays structurally maintainable but out of default query, compile, research, collect, output, and maintenance context. Raw sources become synthesized articles with cross-references and confidence scores. Every directory has an index.md — nothing is scanned blindly. Fast, read-only, index-first answers with exact file citations and bounded searches. Claude uses /wiki:query , Codex gets explicit $wiki-query , and Pi/DS4 get isolated read-only launchers. The full workflow stays available for research and writes. Default-on hook capture writes redacted events, state JSON, and Markdown digests under .sessions/ . Rehydrate future turns with compact context; promote only what belongs in a topic. Curates high-signal corrections, preferences, approvals, and plan acceptance under .sessions/feedback/ . Generic acknowledgements are ignored; durable lessons are explicitly promoted. Score every article for staleness and quality. Two-tier scan: fast metadata check, then deep content read for flagged articles. Checkpoint recovery. Machine-readable JSON + human-readable report. Answer the broader trust question. Reuse the librarian pass, trace outputs across raw/ , wiki/ , and output/ , detect drift, inspect provenance, and do fresh research when local evidence is not enough. Extract lessons learned from the current session — error→fix patterns, user corrections, discoveries. Saved as structured notes the wiki can query later. --rules emits enforceable rules instead of prose. Wiki-grounded implementation plans. Reads the knowledge base, interviews you about requirements, fills gaps with targeted research, and produces a phased plan citing wiki articles as evidence. --format rfc|adr|spec . Reports, slide decks, study guides, playbooks, implementation plans, timelines, glossaries, comparisons. Filed back into the wiki so the next output builds on every previous one. Native plugin. Recommended. claude plugin install wiki@llm-wiki Use /wiki:query for the compact read-only path. Research, ingest, compile, and maintenance commands keep the full workflow. Restart Claude Code after install or update. Marketplace plugin. Invoke with @wiki . codex plugin marketplace add nvk/llm-wiki codex plugin add wiki@llm-wiki Full workflows: @wiki Read-only lookup: type $ and select wiki-query Or from a local checkout: ./scripts/bootstrap-codex-plugin.sh --scope user --verify . The explicit-only $wiki-query skill is about 76% smaller than loading the full workflow for a lookup. Instruction file. In opencode.json: { "instructions": "path/to/llm-wiki/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md" } Query-only alternative: { "instructions": "path/to/llm-wiki/plugins/llm-wiki-opencode/skills/wiki-query/SKILL.md" } The query-only instruction is about 89% smaller and forbids wiki writes. OpenCode package sync and size are tested; live behavior remains provider/model dependent. Full research also requires OPENCODE ENABLE EXA=1 . Read-only launchers. Best for local models. ./scripts/pi-wiki-query current Pi provider ./scripts/pi-ds4-wiki-query isolated local DS4 Full write-capable workflow: pi --skill path/to/llm-wiki/plugins/\ llm-wiki-opencode/skills/wiki-manager/SKILL.md The query launchers expose only read,grep,find,ls , disable unrelated discovery, and load the shared 2.8 KB query profile. Portable instruction files. Small read-only query profile: curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/profiles/query-lite/SKILL.md \ ~/your-project/AGENTS.md Full research and maintenance protocol: curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/AGENTS.md \ ~/your-project/AGENTS.md The query profile is about 95% smaller than the complete portable protocol. Choose the full file only when the agent must write or research. Claude Code: claude plugin update wiki@llm-wiki Restart Claude Code to apply If the update command misses a new version stale marketplace cache , sync manually: git clone https://github.com/nvk/llm-wiki.git or: git -C ~/llm-wiki pull REPO=~/llm-wiki/claude-plugin DEST=~/.claude/plugins/cache/llm-wiki/wiki VERSION=$ grep '"version"' "$REPO/.claude-plugin/plugin.json" | grep -o ' 0-9. ' rm -rf "$DEST"/ mkdir -p "$DEST/$VERSION" cp -R "$REPO/.claude-plugin" "$REPO/commands" "$REPO/skills" "$DEST/$VERSION/" Codex: run codex plugin marketplace upgrade llm-wiki , then codex plugin add wiki@llm-wiki . For a local checkout: re-run ./scripts/bootstrap-codex-plugin.sh --scope user --verify . AGENTS.md: re-run the curl command above to replace the file. One command, from anywhere — creates a topic wiki, launches parallel agents, keeps researching for an hour, comes back compiled. /wiki:research "gut microbiome" --new-topic --min-time 1h More common flows: /wiki:research "nutrition" --new-topic /wiki:research "fasting" --deep --min-time 2h /wiki:research "What makes articles go viral?" --new-topic /wiki:research --mode thesis "fiber reduces neuroinflammation via SCFAs" /wiki:query "How does fiber affect mood?" /wiki:query --resume /wiki add https://example.com/article fuzzy router → ingest /wiki what do we know about CRISPR? fuzzy router → query /wiki I have an idea for cited podcast search capture → inventory/ideas/ /wiki poke holes in the podcast idea shape; no Project yet /wiki use the narrow version and make it a project approve → frozen brief /wiki:ingest-collection https://github.com/bitcoin/bips --wiki bitcoin /wiki:adapter list /wiki:adapter doctor my-private-adapter /wiki:adapter run my-private-adapter --request /absolute/request.json /wiki:collect "bitcoin memes" --wiki bitcoin /wiki:collect "bitcoin memes" --scale medium --media reference --inventory corpus /wiki:inventory add item "TRX-4M ring and pinion" --wiki trx4m-1-18 /wiki:inventory list --view actions --limit 10 /wiki:dataset add "Bitcointalk Temporal Graph" --location https://figshare.com/articles/dataset/BitcoinTemporalGraph/26305093 /wiki:dataset list --view schema --limit 10 /wiki:archive topic old-interest --reason "No longer active" /wiki:archive list --archived /wiki:archive restore old-interest /wiki:session status /wiki:session rehydrate --cwd "$PWD" /wiki:feedback list --unpromoted /wiki:session disable optional opt-out /wiki:compile llm-wiki schema status /path/to/wiki llm-wiki schema adopt /path/to/wiki add advisory topic guide to older wikis /wiki:output report --topic gut-brain /wiki:assess /path/to/my-app --wiki nutrition /wiki:lint --fix Long agent runs lose useful context when chats compact, terminals close, or work moves from Claude to Codex. LLM Wiki now has a hidden operational layer for that: redacted checkpoints, compact digests, and user-feedback candidates under HUB/.sessions/ . It helps the next turn get oriented without treating private chat logs as curated wiki knowledge. A session ID such as codex:abc123 combines the harness name with that runtime's native session ID. It is a local lookup key that ties events, state, and a digest together—not a secret, transcript, wiki, or topic ID. Trusted hooks capture in balanced mode by default. /wiki:session disable writes enabled: false so those hooks become no-ops. Capture stores metadata, state, small events, and distilled Markdown digests. Full transcripts are not stored by default. /wiki:session rehydrate returns a compact context block pointing the agent to relevant digests for the current directory, topic, or session ID. Session digests and feedback candidates remain operational memory until promote writes a distilled note into a topic's raw/notes/ . HUB/.sessions/ ├── config.json enabled/mode/privacy/rehydrate settings ├── registry.jsonl append-only lifecycle events ├── queue/YYYY-MM-DD.jsonl small redacted hook events ├── state/<harness /<session id .json latest per-session machine state ├── digests/YYYY/MM/<harness -<session id .md ├── feedback/ │ ├── candidates.jsonl corrections, preferences, approvals, plan acceptance │ └── status.json └── indexes/ derived caches for cwd/topic/session lookups /wiki:session status /wiki:session rehydrate --cwd "$PWD" /wiki:session list --limit 10 /wiki:session show codex:abc123 /wiki:session promote codex:abc123 --topic meta-llm-wiki /wiki:feedback list --unpromoted /wiki:feedback show fb-abc123 /wiki:feedback promote fb-abc123 --topic meta-llm-wiki /wiki:session disable opt out /wiki:session enable turn capture back on The same .sessions/ layout works across Codex, Claude Code, OpenCode, Gemini-style adapters, and local .wiki/ project wikis. It is separate from the research crash-recovery files .research-session.json and .thesis-session.json and durable workflow provenance .session-events.jsonl and .session-checkpoint.json stored in a topic wiki. .sessions/ ; no full transcripts by default. --fix auto-repairs safe issues. ~/wiki/ Hub — lightweight, no content ├── wikis.json Registry of all topic wikis ├── index.md Lists topic wikis with stats ├── log.md Global activity log ├── .sessions/ Operational session memory, digests + feedback └── topics/ Each topic is an isolated wiki ├── nutrition/ │ ├── .obsidian/ Obsidian vault config │ ├── inbox/ Drop zone for this topic │ ├── inventory/ Items, candidates, corpora, Ideas, views │ │ └── ideas/ Proposals being researched and shaped │ ├── datasets/ Manifests for large/external data │ ├── raw/ Immutable sources │ ├── wiki/ Compiled articles │ │ ├── concepts/ │ │ ├── topics/ │ │ └── references/ │ ├── output/ Generated artifacts │ ├── index.md │ ├── config.md │ └── log.md ├── woodworking/ Another topic wiki └── .archive/ Archived topic wikis, hidden by default Each research area is isolated. No cross-topic noise. Queries stay focused. A multi-wiki peek finds overlap when relevant. wikilinks for Obsidian plus standard markdown links for everything else. Works in every viewer — including no viewer at all. Once a source is ingested it is never modified. Articles synthesize on top. Retraction removes both cleanly. /wiki:collect records aliases, source context, media URLs, cached asset paths, hashes, dedupe notes, scale, and inventory recommendations before anything becomes evidence or durable state. Parts, source queues, corpora, watch items, and next actions live under inventory/ so they can be listed and revisited without becoming evidence. Ideas keep proposal history and research lineage. Promotion requires an approved decision snapshot; the linked Project owns plans, implementation, artifacts, and outcomes. /wiki:portfolio cross-references distributed Ideas and direct or promoted Projects live. Concepts remain supporting knowledge, and no catch-all store can drift. Adapter registrations live in a mode- 0600 machine-local registry. Code and bulk output stay outside the hub; only reviewed wiki-safe candidates may enter normal provenance workflows. datasets/ stores manifests, samples, profiles, and query recipes for large data. The wiki indexes data without copying it into the source corpus. Archived topics live under topics/.archive/ . Most tools skip them by default; deep queries may surface index hits, and explicit --include-archived can read them. .sessions/ is hidden, cross-topic operational state for redacted checkpoints, digests, rehydration indexes, and feedback candidates. Topic promotion is explicit. Runs entirely on the host agent's built-in tools. Plugin is Markdown + commands. No servers, no services, no telemetry. All commands accept --wiki <name to target a topic wiki and --local for the project wiki. Archived topic wikis are skipped by default; commands that support --include-archived require that explicit flag before reading or writing archived material. query , output , and plan also accept --with <wiki for cross-wiki context. | Command | Description | |---|---| /wiki <natural language | Fuzzy intent router — say what you want, including Idea capture, development, shaping, and approved Project promotion. | /wiki | Show status, stats, and list all topic wikis. | /wiki init <name | Create a topic wiki at ~/wiki/topics/<name / . | /wiki:idea new|list|show | Capture Ideas under inventory/ideas/ or browse the active topic and hub-wide catalog. | /wiki:idea develop|shape|promote|archive <slug | Research and challenge an Idea, choose a version, explicitly promote an approved frozen brief into a linked Project, or archive only the Idea. | /wiki:portfolio filters | Show one live, read-only portfolio of Ideas and active Projects across topic wikis without copying records into a catch-all topic. | /wiki:ingest <source | Ingest a URL, file path, or quoted text. | /wiki:ingest --inbox | Process files dropped into the topic wiki's inbox. | /wiki:ingest-collection <source | Bulk-ingest Git doc repos, BIP-style proposal sets, MediaWiki dumps/API sites, message archives, or Wayback CDX snapshots. | /wiki:adapter add <local-path | Register an existing explicitly trusted private adapter with scoped input and output roots; never clone or publish it. | /wiki:adapter list|show|doctor | Inspect machine-local registrations and verify the manifest, executable, and describe handshake. | /wiki:adapter run <id --request <json | Execute a scoped llm-wiki-adapter/v1 request and verify returned artifact paths and SHA-256 hashes without automatic wiki import. | /wiki:adapter remove <id --yes | Remove only the machine-local registration, not adapter code, input, or results. | /wiki:collect "<things " | Find, dedupe, and catalog artifacts, examples, resources, media, memes, tools, entities, or source candidates. | /wiki:collect "<things " --scale tiny|small|medium|large|huge | Control write behavior by operational scale: row count, media weight, provenance burden, volatility, and inventory usefulness. | /wiki:collect "<things " --media archive|thumbnail|reference | Download bounded originals by default for media-bearing collections; use thumbnails for previews or reference to opt out. | /wiki:collect "<things " --inventory corpus|records | Create one corpus record for medium or media-heavy sets, or per-item records only for small durable lists. | /wiki:inventory list | List durable tracking records as compact chat tables or bullets. | /wiki:inventory add <kind "title" | Add items, source candidates, entities, corpora, questions, tasks, or watch records after a fit check. | /wiki:inventory scan-outputs --dry-run | Preview legacy queues/backlogs before an additive inventory migration. | /wiki:dataset list | List dataset manifests without opening samples, profiles, queries, or underlying data. | /wiki:dataset add "title" --location <path-or-url | Index large or external data with a manifest; the data stays outside the wiki. | /wiki:dataset profile <slug --dry-run | Preview lightweight profiling of size, format, headers, or schema observations. | /wiki:archive list --archived | List active topic wikis and optionally archived topic wikis. | /wiki:archive topic <slug --reason "why" | Move a topic wiki to topics/.archive/<slug and hide it from default context. | /wiki:archive restore <slug | Restore an archived topic wiki to active status. | /wiki:archive peek <query | Search archived topic indexes without reading archived article bodies. | /wiki:session status|disable|enable | Inspect, opt out of, or reconfigure default-on redacted session capture. | /wiki:session capture|list|show | Create, browse, and inspect redacted session digests kept outside topic evidence. | /wiki:session rehydrate|promote | Load compact session context or explicitly promote a distilled digest into topic raw notes. | /wiki:feedback list|show | Review redacted feedback candidates for corrections, preferences, approvals, and plan acceptance. | /wiki:feedback capture|promote | Manually add a feedback candidate or explicitly promote one into topic raw notes. | /wiki:compile | Compile new sources into wiki articles. | /wiki:query <question | Compact read-only Q&A with exact file citations. --quick / --deep / --list / --resume / --include-archived . | /wiki:ll | Extract lessons from current session into wiki. --dry-run , --rules . | /wiki:research <topic | 5 parallel agents. --plan multi-path , --deep 8 , --retardmax 10 , --new-topic , --min-time 1h . | /wiki:research --mode thesis <claim | Thesis-driven research: for + against → verdict. | /wiki:plan <goal | Wiki-grounded implementation plan. --format rfc|adr|spec . | /wiki:output <type | summary, report, study-guide, slides, timeline, glossary, comparison. | /wiki:assess <path | Assess a repo against wiki + market. Gap analysis. | /wiki:audit | Truth-seeking audit across wiki, outputs, provenance, and fresh research when needed. | /wiki:librarian | Score articles for staleness and quality; optional conventions pass writes topic-guide proposals without rewriting schema.md . | /wiki:lint | Health checks. --fix auto-repairs and can add a missing advisory topic guide; --deep web-verifies facts. | /wiki:retract <source-path | Dry-run-first removal of a source and its downstream wiki references, including selected archives and sessions. | llm-wiki retract --everywhere --apply | Clean up an accidentally saved sensitive value using hidden or stdin input, common encoded-form matching, explicit application, and local verification. | /wiki:project | Create and manage delivery folders directly or from promoted Ideas. The Project owns changing plans, artifacts, and outcomes. | From zero to a compiled wiki in 5 minutes. 1. Install the plugin claude plugin install wiki@llm-wiki 2. Create a topic wiki Pick any topic you're curious about: /wiki init nutrition This creates a hub at ~/wiki/ and your first topic wiki at ~/wiki/topics/nutrition/ . 3. Research it /wiki:research "gut microbiome and mental health" --wiki nutrition or just say it naturally: /wiki research gut microbiome and mental health Five parallel agents search the web from different angles academic, technical, applied, news, contrarian , ingest the best sources, and compile them into cross-referenced wiki articles. Takes 2-5 minutes. 4. Ask your wiki a question /wiki:query "how does fiber affect mood?" --wiki nutrition or naturally: /wiki how does fiber affect mood? The wiki answers from its compiled articles with citations. 5. Audit before you trust an output /wiki:audit --wiki nutrition /wiki:audit --artifact output/report-gut-brain.md Audit rechecks the wiki layer, traces the output's evidence chain, flags drift, and will do fresh research if the local corpus is not enough to answer the trust question. What to do next: /wiki:research "topic" --deep — 8 agents instead of 5, adds historical and data angles /wiki:research "topic" --min-time 1h — keep researching in rounds for an hour /wiki:research "topic" --plan — decompose into parallel research paths /wiki:audit --project nutrition-playbook — verify outputs and upstream wiki state together /wiki add https://example.com/article — fuzzy router detects the URL and ingests it /wiki what do we know about CRISPR? — fuzzy router detects the question and queries /wiki:lint --fix — clean up structural issues and add a missing advisory topic guide llm-wiki schema adopt /path/to/wiki — safely add schema.md to an older wikiA typical research session flows through four stages: Stage 1: Ask a question or pick a topic llm-wiki auto-detects whether you're asking a question or naming a topic. Use the direct command or the fuzzy router: /wiki:research "What makes long-form articles go viral?" direct command /wiki research quantum computing fuzzy router — same result /wiki:research --mode thesis "fiber reduces neuroinflammation via SCFAs" thesis → for/against evidence Stage 2: Agents search in parallel 5 agents 8 with --deep , 10 with --retardmax search simultaneously from different angles — 2-3 web searches each, full-content fetch, quality scoring 1-5 . A credibility pass deduplicates before ingestion. Stage 3: Sources are ingested and compiled Top sources are saved to raw/ immutable — never modified after ingestion . Then the compilation pass synthesizes them into wiki articles under wiki/concepts/ , wiki/topics/ , and wiki/references/ with cross-references, confidence scores, and bidirectional links. Stage 4: Gap report and follow-up After each round, you see what's covered, what's still missing, and suggested follow-ups. If 2+ gaps remain, you're offered to close them in parallel: Close gaps? 1. Dose-response curves for wavelength specificity 2. Long-term safety data for daily exposure 3. Device comparison clinical vs consumer panels Enter numbers e.g. 1,2,4 , "all", or "skip": Multi-round research: Add --min-time 2h to keep researching in rounds, each drilling into gaps the previous round found. Add --plan to decompose into parallel paths upfront. The hub ~/wiki/ is just a registry. No content — only wikis.json , index.md , and log.md . All content lives in topic sub-wikis. Topic wikis ~/wiki/topics/<name / are isolated research areas. Each has its own sources, articles, outputs, advisory schema.md topic guide, and Obsidian vault config. Isolation means researching quantum computing can't pollute your nutrition wiki. Raw sources raw/ are immutable. Once a paper, article, or data file is ingested, it's never modified. This is the audit trail — every claim in every article traces back to a source. Wiki articles wiki/ are LLM-compiled syntheses organized into three categories: Archive topics/.archive/ is for whole topic wikis the user no longer wants in normal context. It preserves source history, articles, outputs, and logs while keeping old interests quiet by default. Sessions HUB/.sessions/ are operational memory for redacted harness checkpoints, compact digests, rehydration indexes, and feedback candidates. They are not compiled as topic evidence unless explicitly promoted into raw/notes/ . Inventory inventory/ is for durable operational state: actual items, source candidates, corpora, entities, open questions, tasks, watch items, and next actions. It is intentionally not evidence for factual claims. Ideas inventory/ideas/ are proposals that might become deliverables. They link to Concepts and evidence while preserving the original seed, assumptions, alternatives, scope, risks, success criteria, and decisions. They remain catalog records rather than factual evidence. Projects output/projects/<slug / are committed delivery work. Promotion is explicit: an approved Idea snapshot becomes a frozen BRIEF.md beside WHY.md , and the Project then owns all changing implementation truth. Dataset manifests datasets/ let the wiki index large or external data without copying it into raw/ . Manifests can point to local paths, URLs, archives, samples, profiles, and query recipes. Topic guides schema.md are human-owned convention files, not database schemas. They define local vocabulary, relationship verbs, article subtypes, source boundaries, and adoption notes. New wikis get an advisory starter by default; older wikis remain valid and can adopt one with llm-wiki schema adopt or after a librarian conventions proposal. Articles use dual-link format: wikilink for Obsidian + standard markdown links for everything else. Confidence scores high/medium/low reflect source quality and corroboration. Indexes index.md exist in every directory. They're derived caches — rebuilt automatically from file frontmatter. The agent reads indexes first and never scans blindly. Outputs output/ are generated artifacts: reports, slide outlines, study guides, implementation plans. They're built from wiki articles, so every output compounds on all prior research. Audit walks that full artifact graph. It can trace an output back through the wiki state and raw sources it depended on, then escalate into fresh research when the stored evidence is stale or incomplete. Session capture is default-on operational memory. Trusted Codex, Claude, OpenCode, and Gemini hooks write redacted checkpoints into HUB/.sessions/ or a local .wiki/.sessions/ . The topic wiki stays clean until you promote something. What is captured: harness name, session ID, current directory, git branch/remote, small redacted hook events, per-session state JSON, markdown digests, and feedback candidates. Full transcripts are off by default. HUB/.sessions/ ├── config.json ├── registry.jsonl ├── queue/YYYY-MM-DD.jsonl ├── state/<harness /<session id .json ├── digests/YYYY/MM/<harness -<session id .md ├── feedback/candidates.jsonl └── indexes/ Daily commands: /wiki:session status /wiki:session rehydrate --cwd "$PWD" /wiki:session list --limit 10 /wiki:session show codex:abc123 /wiki:feedback list --unpromoted /wiki:feedback promote fb-abc123 --topic meta-llm-wiki Opt out: /wiki:session disable writes enabled: false in .sessions/config.json , so trusted hooks become no-ops. /wiki:session enable turns it back on. Promotion rule: session digests and feedback candidates are useful for continuity, but they become topic knowledge only after /wiki:session promote or /wiki:feedback promote writes a distilled note under raw/notes/ . Use the repo-owned Pi launcher for fast DS4 wiki lookups. It creates isolated Pi state, pins the local deepseek-v4-flash provider settings, disables unrelated extensions and skills, and exposes only read , grep , find , and ls . cd /path/to/llm-wiki ./scripts/pi-ds4-wiki-query Inspect the exact command without launching Pi: ./scripts/pi-ds4-wiki-query --dry-run Optional overrides: DS4 BASE URL=http://127.0.0.1:8000/v1 \ DS4 MODEL=deepseek-v4-flash \ LLM WIKI PI DS4 STATE DIR="$HOME/.pi-ds4" \ ./scripts/pi-ds4-wiki-query Query-only means query-only: the compact profile cannot ingest, compile, edit, or repair a wiki. Switch to the full wiki-manager skill when the requested workflow must write. Measured harness reduction: the controlled AB/BA benchmark reduced the exact serialized provider payload from 216,640 to 77,338 bytes, a 64.30% reduction, while preserving fixture reads, citations, abstention, and no-mutation checks. OpenCode note: its generated query profile receives static and sync validation, but live behavior remains best effort because the user chooses the provider and model. For broad topics, --plan decomposes your research into independent paths and runs them all in parallel: /wiki:research "red light therapy" --plan --wiki redlight-therapy or naturally: /wiki research red light therapy --plan The agent generates a research plan and asks for confirmation: Research Plan — red light therapy Paths will run in parallel 1. Mechanisms — cytochrome c oxidase, wavelength specificity, dose-response 2. Clinical evidence — RCTs for skin, joints, wounds, cognition 3. Devices — LED vs laser, panel sizing, FDA clearances 4. Criticisms — placebo effects, publication bias, safety concerns Estimated: 4 paths x 5 agents = 20 parallel agents Proceed? y/n/edit On confirmation, all paths launch simultaneously. Each path runs its own 5-agent swarm. Sources are ingested in parallel each path writes unique files , then a single compilation pass sees all sources at once for cross-path synthesis. After the first round, a gap report shows what's still missing. Pick which gaps to close — they launch as another parallel batch: Close gaps? 1. Dose-response curves for wavelength specificity 2. Long-term safety data for daily exposure 3. Device comparison clinical vs consumer Enter numbers e.g. 1,2,4 , "all", or "skip": Combine with other flags: --plan --deep — 8 agents per path instead of 5 --plan --min-time 2h — multiple plan-dispatch-compile cycles over 2 hours --plan --new-topic — create the wiki and research in one shotOnce your wiki has compiled articles, generate deliverables from them: /wiki:output report --topic gut-brain --wiki nutrition or naturally: /wiki write a report on gut-brain axis Output types: summary — concise overview of a topic or the entire wiki report — detailed analysis with citations and evidence study-guide — structured learning material with key concepts and review questions slides — slide deck outline with speaker notes timeline — chronological view of events and developments glossary — term definitions extracted from articles comparison — side-by-side analysis of two or more subjectsOutputs are saved to output/ inside the topic wiki and indexed automatically. Every output builds on all compiled articles, so the more you research, the stronger every output gets. Retardmax mode works here too — /wiki:output report --retardmax ships a rough draft immediately. Iterate later. Cross-wiki context: Use --with to pull knowledge from another wiki into your output: /wiki:output report --wiki nutrition --with article-writing Uses nutrition content + article-writing craft knowledge Projects: Start committed work directly, or incubate uncertainty as an Idea first: /wiki I have an idea for rebuilding the blog around reader jobs /wiki research this idea and compare the smallest useful version /wiki poke holes in it /wiki use the narrow version and make it a project Direct Projects remain valid when incubation adds no value: /wiki:project new rebuild-blog "Rewrite the company blog using wiki research" /wiki:output report --topic gut-brain --project rebuild-blog An Idea lives in inventory/ideas/ . Approved promotion creates output/projects/rebuild-blog/WHY.md plus a frozen BRIEF.md and preserves links in both directions. From then on, the Project owns delivery state and its artifacts stay colocated. Research sessions fetch many URLs. By default, Claude Code asks for approval on each one. To skip these prompts, add WebFetch and WebSearch to your project's allow list: In .claude/settings.local.json : { "permissions": { "allow": "WebFetch", "WebSearch" } } This pre-approves all web fetches and searches. You can also allow specific domains only: WebFetch domain:arxiv.org . If you prefer per-session approval, choose "Always allow" when Claude asks about the first URL — that covers the rest of the session. Pulled live from github.com/nvk/llm-wiki/releases https://github.com/nvk/llm-wiki/releases . Falls back to the version baked into the plugin manifest if the API is unavailable. Private adapter protocol. Register and run explicitly trusted local research tools while their code, credentials, corpora, and bulk results stay outside the wiki. Scoped, verified execution. Manifest handshakes, sanitized environments, path policies, timeouts, private receipts, and SHA-256 artifact checks make runs reviewable. No automatic import. Only reviewed wiki-safe candidates may later enter normal provenance workflows; private and bulk artifacts remain external. Pulled live from .claude-plugin/marketplace.json on the master branch via jsDelivr CDN — no rate limits, always current. LLM-compiled knowledge base with explicitly trusted private adapters, a hub-wide Ideas/Projects portfolio, fuzzy Idea shaping and promotion, compact queries, ingestion, audits, sessions, inventory, datasets, archives, and artifact generation. LLM Wiki is a set of commands and a knowledge model that turns any LLM agent into a research engine and an append-only, Markdown-native wiki. It captures rough Ideas, researches and shapes them against compiled Concepts and evidence, and explicitly promotes approved snapshots into Projects. It also runs parallel research, preserves redacted session memory, ingests and compiles sources, audits trust, answers questions, and generates artifacts. Inspired by Andrej Karpathy's LLM wiki concept https://x.com/karpathy/status/2039805659525644595 . Five install modes: Claude Code native plugin , OpenAI Codex marketplace plugin with full @wiki and explicit $wiki-query , OpenCode full or query-only instruction file , Pi/DS4 read-only launchers or the full skill , and any other LLM agent via a compact query profile or complete portable AGENTS.md . The canonical behavior lives under the Claude source tree. Sync scripts generate Codex, OpenCode, and portable query packages, and deterministic tests reject drift. OpenCode live-model behavior remains best effort because its provider and model are user-selected. By default in ~/wiki/ on your machine. You can relocate the hub to iCloud Drive, Dropbox, or any custom path: /wiki config hub-path ~/Library/Mobile\ Documents/com~apple~CloudDocs/wiki Config lives at ~/.config/llm-wiki/config.json . Shared hubs should use portable hub path values and relative wikis.json topic paths; old machine-specific resolved path values are treated only as fallback caches. macOS privacy permissions are local to each Mac and to the exact process launching the agent. One machine can read the iCloud hub while another machine with the same Apple ID and same files returns Operation not permitted . If stat succeeds for the hub but reading wikis.json or listing topics/ fails, the configured hub path is correct. Grant Full Disk Access or iCloud Drive access to the launcher app, then fully restart that app and the agent. For SSH-launched Codex sessions, Terminal.app permission is not enough because Codex is a child of sshd-session . Grant access to /usr/libexec/sshd-session , /usr/libexec/sshd-keygen-wrapper , and optionally /usr/sbin/sshd , then start a new SSH session. Session memory lives in HUB/.sessions/ or a local .wiki/.sessions/ . It is hidden operational state: redacted hook events, state JSON, digest Markdown, feedback candidates, and indexes for rehydration. It is default-on so long sessions can survive compaction and future turns can start with the right context. It does not store full transcripts by default. Run /wiki:session disable to opt out, /wiki:session rehydrate to load compact context, and /wiki:session promote only when a digest should become topic evidence. schema.md is a human-owned topic guide. It is not a database schema and not an automatic migration target. Use it to capture a topic's local vocabulary, relationship verbs, article subtypes, source boundaries, and convention notes. New topic wikis get an advisory starter by default. Existing wikis without one are still valid; run llm-wiki schema adopt --dry-run /path/to/wiki to preview the starter or llm-wiki schema adopt /path/to/wiki to add it. For established wikis, run the librarian conventions pass first and copy only the useful proposal parts into schema.md . Yes. Each topic wiki ships with its own .obsidian/ vault config and can be opened as an independent vault: open ~/wiki/topics/nutrition/ Cross-references use a dual-link format: gut-brain-axis|Gut-Brain Axis Gut-Brain Axis ../concepts/gut-brain-axis.md Obsidian reads the wikilink for graph view and backlinks; everything else Claude Code, GitHub, plain text editors follows the standard markdown link. A research mode inspired by Elisha Long's retardmaxxing philosophy https://www.retardmaxx.com/ — act first, think later. Ten parallel agents, skip planning, cast the widest net, ingest aggressively, compile fast, lint later. Available on /wiki:research --retardmax and /wiki:output --retardmax for when you want results now and will clean up afterward. /wiki:research --mode thesis "<claim " starts from a specific claim and uses it as a filter. Agents are split across supporting , opposing , mechanistic , meta/review , and adjacent — balanced by design. Sources that don't relate to the claim's variables are skipped, which keeps the wiki tight. Output is a verdict : supported, partially supported, contradicted, insufficient evidence, or mixed. With --min-time , round two focuses harder on the weaker side of the evidence — counter-weight against confirmation bias. Yes. LLM Wiki is MIT-licensed. Source and releases at github.com/nvk/llm-wiki https://github.com/nvk/llm-wiki . Compiling, querying, linting, and generating artifacts from an existing wiki work offline — everything is plain Markdown on your disk. Research and ingestion need internet since they fetch URLs and search the web. Zero runtime dependencies. LLM Wiki uses only the built-in tools of the host agent file read/write, web fetch, web search . The plugin itself is Markdown: command definitions, skills, and reference docs. Optional: ask-grok-mcp https://github.com/nvk/ask-grok-mcp for best-in-class tweet ingestion, tobi/qmd https://github.com/tobi/qmd for local search beyond ~100 articles. Claude Code: claude plugin update wiki@llm-wiki and restart. Codex: run codex plugin marketplace upgrade llm-wiki , then codex plugin add wiki@llm-wiki or re-run the bootstrap helper for a local checkout . AGENTS.md: re-curl from the master branch to replace your file. See Install install for the manual sync fallback when the plugin updater hits stale cache.