Obsidian is the viewer. MemoryOS is the brain.

My LinkedIn DMs have been busy this week with one question.

Mehmood Ferozuddin asked it most directly this morning: "you posted about MemoryOS using a local DB. Why not Obsidian? It also local and can be two-way if configured properly in Claude Code."

I get a version of that question almost every week now. The wave is real, and you have probably seen it.

Andrej Karpathy dropped a gist called the LLM Wiki. The premise: Obsidian as your raw vault, Claude Code as the librarian, a CLAUDE.md schema file as the rulebook. The wiki grows itself. He framed it nicely too: "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase."

Every prompt-engineering influencer ran with it. In the last 30 days I have watched five separate viral threads claiming "Obsidian + Claude Code = your second brain." Capital One assessment teams. BIM engineers wiring Obsidian to Revit. Solo founders calling their vault Janus and treating it like a chief of staff. The pattern is on Reddit, X, LinkedIn, and YouTube. It is the moment.

And it is also pre-2026 thinking dressed in new packaging.

The reframe

It is the same one I sent Mehmood, and probably answers you if the same question has been sitting in your head.

Obsidian is a markdown vault optimized for the human reader. You open it. You scroll. You link notes. You watch the graph view bloom. Plugins add semantic search, daily notes, mobile sync. Every design decision points at the person reading.

MemoryOS is a Postgres brain optimized for the AI reader. Typed knowledge (fact, pattern, insight, lesson). Confidence and decay scoring per row. tsvector plus pgvector for ranked recall in a single query. A separate decisions table that stores options_considered and rationale, so six months later you can query why you chose X over Y. The human is a guest. The AI is the primary user.

Different primary user, different optimization target. That is the whole difference.

The self-admission

The proof that everyone already knows this: look at what the serious memory tools shipped in the last 30 days.

Hermes Agent (Nous Research, 100K GitHub stars, MIT) is a markdown-first architecture. It is also openly broken at scale. Rohit Ghumare, who evangelizes it, wrote the kill quote himself:

"MEMORY md file is 2,200 chars. USER md file is 1,375 chars. Hit 80% and consolidation kicks in: the agent merges related entries into denser versions, which is lossy. The longer you run Hermes, the more your earlier context gets compressed away."

The longer you run it, the more your earlier context gets compressed away. A markdown-only system telling on itself.

The patch-on-top wave

So Hermes added a Tier 3 slot for a structured memory plug-in. That slot is exactly where every other 30-day shipper bolted on:

• agentmemory: BM25 plus vector plus knowledge graph with reciprocal rank fusion, Ebbinghaus decay, 92 percent token savings
• MemOS 2.0: tiered semantic memory with auto scoring
• Zilliz memory library: markdown files as the presentation, vector store rebuilt from them
• Delta-mem (VentureBeat): 0.12 percent parameter add-on that outperforms RAG
• MEMO (MIT, A*STAR, NUS, Liquid AI): memory as a separate model
• Letta Code: memory-forward UI with graph view, commit history, markdown as a viewer over structured state
• Oracle: pitched Oracle AI Database as the persistent memory core
• Mem0: one-command flow for agents to self-onboard to a vector memory layer in five seconds

Every one of them kept markdown as presentation. Every one of them built a database for the substrate.

The cost pressure

Microsoft reportedly scaled back internal Claude Code licenses because token billing got hard to justify. Uber's CTO told the company they burned through the entire 2026 AI budget in four months. Grepping 500 markdown files on every session is the most expensive way to remember. Vector recall and ranked retrieval cut that 14 to 92 percent depending on the benchmark.

The honest part

Obsidian wins on human UX. Graph view. Plugins. Mobile app. Daily notes. The feel of opening a vault and scrolling. If you read your notes, Obsidian belongs in your stack.

I use both. The vault as a viewer, the SQL brain as the substrate. Markdown for me. Postgres for the AI.

If you want the longer breakdown with the schema diffs side by side, I put it here.

And the system itself lives here.

Markdown is a presentation format. SQL is a knowledge substrate. You can render SQL into markdown. You cannot reliably query markdown.