Engineers use markdown wikis as lightweight memory for Codex, Claude Code, and Hermes

• Markdown is becoming the boring substrate for agent memory: LangChain's OpenWiki launch split wikis into code and personal brains, while coreyganim's GitHub note wired one company knowledge base into Codex, Claude Code, and Hermes.
• The common shape is files first: coreyganim's Second Brain thread used raw/, wiki/, outputs/, and CLAUDE.md, while the CMOHQ walkthrough ran a marketing department from folders and markdown.
• The technical stack is moving retrieval behind a serving layer: a GraphRAG retrieval design used graph, deep, and agentic search, and an MCP boundary note kept agents away from direct database access.
• Memory now has maintenance work: one recency-bias report showed polluted memory treating model output as ground truth, while the /checkup post and the Claude Code 2.1.206 changelog added cleanup paths for CLAUDE.md, skills, MCPs, and stale sessions.

LangChain pointed readers to the brains launch blog and the open-source OpenWiki repo. The agent-memory demo scene got wonderfully practical: a Friday cron that refreshes a GitHub knowledge base, a zip-file marketing OS with an installer.md, daily podcasts generated from a personal wiki, and a Claude memory bug that turned its own DNA analogy into fake project history.

OpenWiki added two memory modes: code brain for a codebase wiki and personal brain for broader work like email and web tasks. LangChain's launch post called the new abstraction “brains” and said codebase and personal wikis need distinct configuration and run workflows.

The pitch around open memory was portability, not a smarter chatbot. Vtrivedy10's summary described Brace's system as transparent in how memory is generated, compatible with open models, integrated into existing work systems, markdown-first, and updated over time.

The live question is the shape of memory. one follow-up framed wiki memory as a map of where to search, not a bag of facts, and left markdown hierarchy versus graph traversal as an open design choice.

coreyganim described a company knowledge base that lives in GitHub, connects to Codex, Claude Code, and Hermes, and gets refreshed every Friday by a cron job. That is the cleanest version of the pattern: put the company brain somewhere agents can read, diff, and update.

The earlier Second Brain layout was four pieces:

• raw/: transcripts, notes, screenshots, SOPs.
• wiki/: the organized version written by AI.
• outputs/: answers and reports generated by AI.
• CLAUDE.md: instructions for how the AI organizes the system.

A later reply said the same system can run as one shared brain across businesses or one brain per business in the same thread.

coreyganim's CMOHQ notes make the no-code version concrete. The marketing system runs from folders and markdown files, works with Claude, ChatGPT, Hermes, and Codex, and ships as a zip file.

The mechanics were listed as an operating system, not an app:

1. Folders and markdown replaced n8n, LangChain, and CrewAI for a nontechnical copywriter.
2. One HQ folder stores identity, and each brand has intelligence, departments, and metrics.
3. A finance folder was built from a YouTube transcript and ingests ThriveCart, Kit, Google Ads, and Zeni exports.
4. iCloud stores the files so phone, laptop, and office share the same brain.
5. The publishing pipeline goes from approved article to Ghost via MCP, with ElevenLabs audio.
6. Reporting uses MCPs for Kit, Google Analytics, Microsoft Clarity, and monthly P&L data.
7. Claude Cowork handles the workflow because most of the work is writing.
8. Jobs To Be Done become folders, and folders get instructions.
9. The zip includes an ops manual and an installer.md that interviews the buyer.
10. The two primitives are memory, instructions, and tools, plus a repeatable sales deck.

The interesting part is not the folders. It is that delivery became a portable memory product.

The stronger technical version separates storage from retrieval policy. pauliusztin_'s GraphRAG thread put the whole knowledge graph in one database, then exposed memory through an MCP server instead of letting the harness talk directly to the database.

The retrieval modes were distinct:

• Graph search: run full-text and vector search in parallel, fuse rankings with RRF, traverse 2 to 3 hops, rerank, and keep the top 10.
• Deep search: save a larger retrieved set to disk and build a temporary LLM wiki for progressive disclosure.
• Agentic search: let the model write the database query, then validate syntax and enforce permissions.

The boundary claim was simple: databases store knowledge, MCP owns behavior. one follow-up put search, updates, and protection in the MCP server, with Prefect keeping the knowledge fresh through durable ingestion workflows.

In the larger architecture sketch, the harnesses, Claude Code, Codex, OpenCode, and Gemini CLI, became replaceable clients of a portable context layer.

mattpocockuk's personal wiki had already moved past static notes. He described a weekday 15-minute podcast generated from new wiki inputs across X, Discord, Slack, Gmail, and GitHub issues, plus on-demand topic podcasts before making videos.

A custom CLI lets agents inspect his course material and AI Hero posts from the wiki. In a reply, he said the big updates came from his skills, not from the models in the model-upgrade reply.

Hermes moved in the same direction from the other side: export the agent sessions themselves. Teknium's update said Hermes can export conversations to HTML, Markdown, JSON, and more, or upload selected session datasets to private Hugging Face repos.

The export surface included machine round-trip JSONL, readable Markdown archives, Quarto docs, standalone HTML, prompt-only exports, trace exports, filters by model/date/source/cost/branch, redaction, SHA-256 verification, and lineage stitching across compacted sessions

.

Opaque memory started showing teeth. nptacek's report said Opus 4.8 with memory was overweighting recent chats and treating its own generated text as ground truth, including a “BLAST for loom” idea that reappeared later as if it were implemented work.

The same corpus mining also surfaced upside. another nptacek post showed an agent rediscovering orphaned projects in old chat archives and proposing a structured namespace to stop paying a rediscovery tax on the same 5 to 10 specs.

Claude Code is adding cleanup affordances around the same pressure points. the /checkup post listed unused skills, MCPs, plugins, duplicated CLAUDE.md content, root CLAUDE.md splitting, slow hooks, auto mode, and pre-approved read-only commands as things /checkup can inspect before changing.

The next CLI changelog added a /doctor check that proposes trimming checked-in CLAUDE.md files by removing content Claude can derive from the codebase. The same release fixed stale background-agent upgrades, MCP timeout handling, OAuth MCP re-authentication, and misleading expired-login model errors in the 2.1.206 changelog.