Cognition compare
Cognition vs Obsidian and LLM wiki setups
Obsidian is a vault of Markdown notes you curate and read. An LLM wiki is usually search or RAG over those notes when a chat needs a quote. Cognition is not a lookup layer over static files: it stores approved SKILL.md procedures in Postgres, matches them to live agent intent on configured assist calls, pushes the body into the agent context, and can refuse or ask before apply when context or approval rules say so.
How Cognition works in the product
Not a wiki lookup: behavior shipped in this codebase today.
- On meaningful work the install stub requires get_pedagogy_instructions, then cognition_assist with the user's intent and open file paths so hosted operating rules load before the agent acts.
- When the host calls cognition_assist, assist scores skills against triggers, keywords, paths, and embeddings, then returns full SKILL.md bodies in the MCP tool result so the agent reads them in context immediately.
- Skill saves require approval: cognition_save_skill fails closed unless approved_by_user is true, and the server runs privacy scan, policy lint, duplicate review, and author ownership checks before inserting into clo_skills.
- A skill is a procedure, not a note: saves expect structured sections (When to fire, Steps, Checks) that the matcher and gates can parse.
- Apply-time gates: assist can withhold or require confirmation for cross-user skills, heavy matches, privacy flags, or context-risk mismatches (for example prototype guidance in a regulated project).
- Outcomes close the loop: cognition_report_skill_outcome records whether reuse helped, which feeds teachability and whether the skill surfaces again on similar intent.
Cognition is for
You and your coding agents
Developers who want coding agents (Cursor, Codex, Claude Code, Gemini) to receive approved procedures at runtime via MCP, with explicit save consent, load receipts, and outcome reporting.
Obsidian / LLM wiki is for
Their primary user
People who want a durable note library they browse and link, optionally synced, and optionally queried by search, RAG, or MCP plugins over vault files.
Side by side
| Aspect | Cognition | Obsidian / LLM wiki |
|---|---|---|
| Primary unit | SKILL.md row in clo_skills with trigger columns and executable sections | Markdown files in an Obsidian vault (plus attachments) |
| How the agent gets context | MCP tool results inject matched skill bodies into the session | You or the agent query the vault; chunks or links are retrieved on demand |
| Write path | Agent drafts from work; server refuses save without approved_by_user | You edit notes directly; no built-in agent approval gate |
| Match model | Intent plus paths scored against triggers and keywords; can abstain or ask | Similarity or keyword search over note text (in a wiki/RAG setup) |
| Wrong-context protection | Context-risk gate can block or flag skills by posture (prototype vs regulated) | Depends on your wiki pipeline; not intrinsic to Obsidian |
| After reuse | Outcome tool updates learning state on the skill row | Note history in git or Obsidian; no skill outcome loop unless you build one |
When to pick Obsidian / LLM wiki
- You want a personal knowledge base to read and link, not agent procedure memory.
- Your stack is vault plus embeddings plus chat, and you do not need approval-gated skills.
- You need offline-first local files only, with no hosted MCP service.
How they work together
Keep Obsidian for human-readable notes and long-form reference. Use Cognition when a workflow should become an approved skill that agents load automatically on matching intent, not when someone remembers to search the vault.
Try Cognition free for two weeks
Memory and judgment for you and your agents. Install in about two minutes for Cursor, Codex, Claude Code, or Gemini. Then $9.99/mo, cancel anytime.