An open spec for an AI-era wiki.

Every AI dev tool — Claude Code, Cursor, Codex, Gemini CLI, Windsurf, Aider — wants per-project context. Today users hand-curate that context in CLAUDE.md / AGENTS.md / .cursor/rules and it goes stale within a week. memori.wiki replaces the curation step with an AI-readable URL that returns a structured representation of a person's knowledge graph. Any tool that follows this spec can fetch the same payload and stay in sync.

The spec is intentionally minimal. Markdown for content, URLs for addressability, query parameters for token-economy knobs. No bespoke schema, no SDK lock-in. This is on purpose — the only way the spec earns adoption is by being cheap to implement.

Three scopes, each addressable, each AI-readable.

During the rebrand transition, memory.wiki/{id}, memory.wiki/b/{id}, andmemory.wiki/hub/{slug} resolve to the same payloads. mori.wiki/... 301s to memori.wiki/....

Every public URL has a clean-markdown sibling at /raw/...:

GET /raw/{docId}
GET /raw/b/{bundleId}
GET /raw/hub/{slug}
GET /raw/hub/{slug}/c/{concept}

Response is text/markdown, charset UTF-8, with a YAML frontmatter block telling the AI which scope it just received. Browsers viewing these URLs see plain markdown.

The middleware also routes {id}.md / /b/{id}.md / /hub/{slug}.md as convenience aliases so users can paste a .md suffix and get the raw form directly.

Three knobs control token economy on raw responses:

Knobs combine. ?compact&full=1 returns every doc inline, whitespace stripped — useful for one-shot full-bundle context. ?compact&graph=0 returns the doc inventory only, no analysis.

Following the llms.txt standard: every public hub publishes a crawl manifest at:

GET /hub/{slug}/llms.txt

The body lists every public doc and bundle with title, URL, and ≤200-char description, plus pointers to the?compact and ?digest variants. Agents that prefer a manifest crawl over a one-shot payload fetch llms.txt first, then walk to specific docs on demand.

/hub/{slug}/llms-full.txt serves a denser variant capped at 80k tokens by default (override with ?cap=) for AI tools that prefer one fetch over many.

Bundle responses ship the canvas analysis inline so the receiving AI inherits the prior AI's work for free:

---
mw_bundle: 1
id: <bundleId>
title: "..."
url: https://memory.wiki/b/<id>
document_count: N
updated: <ISO>
analysis_generated_at: <ISO>
analysis_stale: true        # only when a member doc was edited after the analysis
source: "Memory.Wiki"
---

# <Bundle title>
> <description>
**Intent:** <intent>

## Summary
<canvas summary>

## Themes
- ...

## Cross-document insights
- ...

## Key takeaways
- ...

## Open questions / gaps
- ...

## Notable connections
- **doc A** ↔ **doc B** — <relationship>

## Concepts (this bundle)
- **concept** (from **doc title**)

## Concept relations
- **conceptA** ↔ **conceptB** — <edge label>

1. [Doc 1](https://memory.wiki/<docId>) — annotation
2. [Doc 2](https://memory.wiki/<docId>) — annotation
...

One endpoint, hybrid retrieval (semantic + keyword + optional rerank). Hub-scoped, public.

POST /api/hub/{slug}/recall
Content-Type: application/json

{
  "question": "react hooks vs context",
  "k": 5,
  "level": "doc",         // doc | chunk | bundle
  "hybrid": true,         // chunk-level only — BM25 + vector reciprocal rank fusion
  "rerank": true          // Anthropic Haiku cross-encoder; doubles latency, raises precision
}

{
  "hub": { "slug": "you", "display_name": "..." },
  "results": [
    {
      "id": "abc123",
      "title": "React hooks crash course",
      "url": "https://memory.wiki/abc123",
      "snippet": "...react hooks let you...",
      "distance": 0.13,
      "updated_at": "2026-04-10T..."
    }
  ]
}

k capped at 20. distance is cosine — lower is more relevant.level=chunk returns paragraph-level matches with doc id + heading path.level=bundle matches against bundle titles + descriptions.

The hub-wide concept index is the AI-era replacement for hand-typed categories and tags. Built incrementally by the bundle analysis pipeline — every Analyze run upserts concepts the LLM extracted from that bundle's docs.

Public surface:

GET /hub/{slug}/c/{concept}           # rendered per-concept page (HTML)
GET /raw/hub/{slug}/c/{concept}        # same, raw markdown variant

Per-concept page returns: canonical label, description, source documents, neighbour concepts. The neighbours come from concept_relations typed edges (supports / elaborates / contradicts / exemplifies / contains). Same edge vocabulary the bundle canvas uses.

Three states, applied uniformly to docs, bundles, and hubs.

Every gating decision happens server-side in the raw-fetch routes — there's no path by which a URL can leak content the rendered viewer wouldn't already show. Revoking allowed_emails takes effect on the next fetch.

Traditional wikis (Wikipedia, Roam, Obsidian) build their graph by having humans type[[other page]]. memori.wiki doesn't ship that syntax. It's a deliberate omission.

[[wikilink]] is a 2003-era mechanism for compensating for AI's absence. A user typing[[React Hooks]] tells the system “these pages are related.” In 2026 the AI does that work without being asked: concept_index extracts "React Hooks" as a concept from any doc that meaningfully discusses it, and concept_relations wires those concepts to neighbours by typed edges.

The user-facing replacement is the Related in this hub panel rendered below every public doc — other docs that share concepts with the current one, surfaced by the AI's analysis. Same information traditional backlinks would give, without anyone having to type the link.

The AI does the linking; you write.

Three integration paths, all stable.

1. URL paste. Drop a memori.wiki URL into any AI chat. The model fetches the raw variant automatically. Zero config, works today across Claude / ChatGPT / Cursor / Gemini.
2. Context file reference. Add the URL to AGENTS.md / CLAUDE.md /.cursor/rules. The AI dev tool fetches on every session boot. See /docs/integrate for per-tool snippets.
3. API integration. Call POST /api/hub/{slug}/recall directly for precise retrieval. No SDK required — plain HTTP. The endpoint is hub-public, no API key needed.

The reference renderer is markdown-it with footnotes + GFM, configured identically to TipTap so authoring and viewing surfaces produce the same DOM. Source: github.com/raymindai/mdcore.