Mythos

Enrichment is a set of 📝Augmentation controls that make a MythOS library legible to AI agents — link suggestions that surface as you write, a write-time link lint that keeps the knowledge graph connected when agents write, one-line memo descriptions agents scan before opening anything, and a health check for semantic search. It exists because agents drafting at speed skip the linking step a careful human takes: they save memos naming people, companies, and concepts the library already holds without linking a single one, and every unlinked write quietly starves the graph, backlinks, and retrieval. Enrichment is for creators who work with AI agents inside their library — especially over the MythOS MCP — and want every write to strengthen the corpus instead of fragmenting it.

Key Capabilities

  • Link suggestions as you write — in a memo or the Journal, titles your library recognizes get a subtle dotted underline; hover for a one-click Insert, or dismiss it for that document. An Enrich All button at the top of the memo adds every pending suggestion at once, after confirming how many memos it will link — each linked once, at its first occurrence. Collaborators editing a memo get suggestions too, drawn from their own library. Always a suggestion, never a block.
  • Write-time link lint — every memo write is checked for known titles it names without linking; off, suggest (default), or require.
  • Require mode for agents — agent writes introducing unlinked references are rejected until linked or explicitly acknowledged; human edits are never blocked.
  • You choose which names your agents must link — a memo named with an everyday word like work or trust is not required by default, because those words appear in ordinary writing without referring to your memo. Names that are distinctly yours are. Review the list at Settings → Enrichment and mark any name required or not required — that is how a memo named for a real company or product, whose name also happens to be an everyday word, gets treated as the entity it is.
  • Aliases and opt-outs — alternate names a memo is recognized by, plus a per-memo opt-out that removes a memo from recognition entirely.
  • AI memo descriptions — one-line agent-facing summaries generated with Claude Haiku on your own Anthropic key, cost-estimated before anything runs.
  • Semantic search health — see how much of your library is indexed for search and chat, and reindex whatever is missing or out of date.

Getting Started

Everything manages from one settings page: Settings → Enrichment (/settings/enrichment).

  1. Choose how strictly your AI agents must link the memos they mention — Off, Suggest (default), or Require.
  2. Before switching to Require, review which names are required. The same page lists every name your agents would be asked to link, split into required and not required, with the reason for each. Everyday words are excluded for you; promote any that are really a company or a product, and demote anything that would be noise. It takes one pass, and after it Require is safe to leave on.
  3. Write normally — type-time suggestions are on by default, so recognized titles get a dotted underline in memos and the Journal; hover one and click Insert to turn the mention into a link.
  4. Add an Anthropic API key under Settings → AI, then generate descriptions across your qualifying memos — the estimated cost shows before anything runs.
  5. Tune any single memo — edit its description, add the other names it goes by, or change whether its name is required.
  6. Check search coverage and reindex whatever is missing or out of date, with the cost shown first.

FAQ

  • What does require mode actually block? Only agent and API writes that introduce a new unlinked reference to a known memo. Pre-existing unlinked names are reported but never block, and human edits in the editor always stay suggestion-only.
  • Won't Require reject everything, if I have memos named after common words? No. A memo whose name is an everyday word is not required by default — the flood that would cause is the reason the distinction exists. Only names that genuinely identify something get enforced, so a rejection always means the agent skipped a real reference.
  • Can an agent override a rejection? For a missing link, yes — deliberately. The rejection lists the unlinked memos, and the agent either rewrites the content with @mentions or retries while acknowledging specific memo IDs. Forgetting cannot produce an override; only an explicit act can.
  • Is there anything an agent cannot override? Yes, and it is the important exception. If a write would expose a private memo — a public memo mentioning one that isn't public — that is a privacy boundary, not a linking suggestion. It cannot be acknowledged away at any setting, and it is blocked even when the writer has no access to the memo in question. The only fix is to remove the mention.
  • What do descriptions cost? They run on your own Anthropic key, never platform billing — fractions of a cent per memo. A 12,000-memo library filtered to roughly 7,000 qualifying memos runs about $10, and the dashboard shows the estimate before you confirm.
  • Will regeneration overwrite a description I wrote by hand? No. Hand-edited descriptions are never touched by the bulk pass. The only way to replace one is the per-memo Regenerate button, which asks first.
  • Which memos qualify? Memos with at least 200 characters of content and at least one tag or link. The same rule governs suggestions, lint matching, and description generation, so short stubs never create noise or generation cost. Qualifying is not the same as being required, though: whether a memo is substantial enough to deserve links and whether its name is safe to enforce are separate questions, which is why a long, well-written memo named work still gets a description and still gets suggested, but never blocks a write.

Related

Contexts

Created with 💜 by One Inc | Copyright 2026