A second brain is a personal library of saved knowledge: articles, PDFs, transcripts, highlights, notes, and decisions kept so those materials can be reused later. The folder or app that holds them is often called a vault. Capture is the intake step that moves a page, PDF, or highlight into that vault. An agent is a system that can read and edit files, run tools, and keep working across turns. A vault constitution is a short, always-loaded instruction file (often named CLAUDE.md or AGENTS.md) that states the vault map, naming rules, and non-negotiables for any agent that touches the files. A promote gate is the review step that decides whether a compiled draft may leave the inbox or raw sources layer and enter the lasting wiki layer.
Capture tools filled the pipe years ago. Chat prompts did not become a durable operating system for that pipe. The 2026 practice that compounds seats agents inside a local Markdown vault under a short constitution and only promotes compiled notes from inbox into wiki after a gate. Continuous chat overwrite of live notes is the common failure mode that looks productive and quietly ruins the archive.
What people were optimizing instead
Operators building second brains chased two incomplete scoreboards. The first was capture volume: more clips, more highlights, more daily dumps into Pocket, Matter, Readwise, and similar front doors. Intake got cheap. The unread pile and the highlight warehouse got large. The second was conversational memory: longer system prompts, sticky chat memories, and "remember this about me" features inside hosted products. Session continuity improved. Ownership of an editable evidence corpus did not.
Local Markdown vaults such as Obsidian made a third path available. Humans and agents could read and write the same files. Linking was native. The product surface that sold the category was often the graph view or the chat pane sitting on top of notes. Neither surface is an operating system for agent write. A graph does not decide what may become evergreen. A chat turn does not persist routing rules when the next agent session starts cold.
Mechanism gallery: the OS that shows up in stacks that compound
Several mechanisms now appear together in second-brain setups that survive reuse.
Constitution before chat. Serious stacks put a short constitution at the vault root. The file holds architecture rules, folder map, naming conventions, and hard bans (do not rewrite raw sources; cite originals; prefer append over overwrite). Procedures live in modular skills (SKILL.md files and similar) so the constitution stays short enough that an agent actually follows it. Updating the constitution is how operators "train" the system. Re-correcting the same mistake inside a chat does not stick across sessions.
Inbox and raw stay append-only. Capture still lands in an inbox or raw layer: web clips, transcripts, voice dumps, conference photos, Readwise-style highlight exports. That layer is treated as near-read-only after ingest. A compiled wiki or notes layer holds structured claims that cite those originals. Agents may grow the compiled layer. They should not silently edit the evidence layer into the model's paraphrase of yesterday.
Draft off to the side, then promote. Draft edits often happen in a branch, worktree, or staging folder rather than against live gold notes. A critic pass (human, subagent, or both) checks the diff for broken links, missing provenance, contradictions, and bulk rewrite risk. Only then does a promote step move material from inbox into wiki. The vault grows by append and promotion. It does not grow by overnight overwrite of everything the agent can see.
Specialized overnight crews, not one omniscient chat. Operators who batch work overnight split roles: ingest and clean messy capture, file inbox items, propose links, run research with citations, and guard against dangerous writes. Overnight cost can run higher than a single prompt. The bet is compounding state in files, not a prettier chat transcript by morning.
Human procedures first, then automation. Writeups that stick often reverse the popular order. Operators define linking schemes, frontmatter templates, and folder scoping themselves until the rules are boring. Only then do they encode those rules into a constitution and skills for an agent with CLI access. Teaching an agent a half-specified vault recreates the unpredictability people blame on the model.
Capture and research-tooling inventory
Capture remains abundant. Highlight hubs syncing into Obsidian stay popular and largely append-only. Mobile widgets, shortcuts, and voice dumps feed the same inbox. Product names here are existence proofs of demand for cheap intake. They are not a ranking of vendors and not a claim about user counts.
What changed is where scarce attention goes after intake. Research tooling and vault agents are being asked to compile, cite, and stage writes. Coding-agent CLIs with vault skills, open agent-skills repos aimed at Obsidian Markdown and search, and sample AGENTS.md frameworks for wiki maintenance are existence proofs of demand for constitution-plus-gate control. Local full-text indexes built beside the vault show up when embedding-only chat retrieval feels too slow or too lossy for daily work. Session-start playbooks (reindex, load constitution, resume open tasks) turn the vault into durable state rather than a bag of notes the agent rediscovers each morning.
Automated research stacks that keep Plan / Search / Read / Extract / Verify / Cite as phases fit the same OS. Verification and citation are promote-adjacent steps. A single-stage summary dump into the wiki without a gate is capture wearing a research costume.
Objections
"A longer system prompt is enough." Prompts evaporate when the product, model, or session changes. A constitution file lives next to the notes, travels with the vault, and can be reviewed in git. Prompts are convenient. Constitutions are portable governance.
"Auto-promote overnight is fine if the model is good." Unattended promote recreates write-access rot: confident prose replaces originals, duplicates fork, and provenance disappears. The scarce control is the gate, not another point of model quality. Overnight crews that stage drafts for morning review keep the speed of batch work without handing gold memory to an unsupervised rewrite.
"Folders and PARA-style taxonomy are already an OS." Taxonomy helps humans browse. It does not, by itself, constrain agent write. Folder scoping can sandbox experiments. Without a promote rule, agents still treat every Markdown file as fair game. Taxonomy plus gate is an OS. Taxonomy alone is a map.
"Chat product memory replaces the vault OS." Hosted memory stores snippets for continuity. It is not an owned inbox-to-wiki pipeline an operator can edit, version, delete, and reopen offline. Snippet memory evaporates when the product changes. A vault with a constitution and a promote gate remains inspectable.
Implications
The market is moving from second brains scored on capture volume or chat fluency toward stacks scored on whether agent write is governed. Constitution files, skill packs, staging folders, critic passes, and promote rituals all point at the same demand: compounding knowledge that stays checkable after the agent leaves.
For product design, Follow / Send / Ask maps onto this constraint set as infrastructure. Follow selects sources worth capturing. Send delivers finite digests into an inbox the agent can read without rewriting originals. Ask queries a wiki that only contains material that passed a gate. Drop the gate and Ask returns whatever last night's model believed. Keep the gate and Ask returns an owned corpus.
Decision criteria for any second-brain or research stack with agents:
- Is there a short, file-based constitution the agent must load, separate from disposable chat prompts?
- Is there an append-only inbox or raw layer that agents do not rewrite in place?
- Do drafts stage off to the side (branch, worktree, or staging folder) before they touch gold notes?
- Is there an explicit promote step with a critic check for provenance, links, and contradictions?
- Can an operator point to what entered the wiki yesterday and open the originals behind it?
Stacks that fail those tests still look finished: busy capture, fluent chat, dense graphs. Stacks that pass them compound because every overnight run leaves governed state rather than a rewritten past.