Memory.Wiki roadmap

A snapshot of what's shipping, what's queued, and what's deliberately not on the list.

Live now (v6)

  • URL primitive at three scopes: Document, Bundle, Hub, all fetchable as markdown.
  • Capture from anywhere: paste, file drop, ChatGPT/Claude/Gemini share URLs, /memory.wiki in Claude Code / Cursor / Codex / Aider, Chrome extension, MCP server.
  • Memory.Wiki Memory pipeline: doc, chunk, bundle embeddings (idempotent), public Recall API with vector and hybrid BM25 plus RRF, hub semantic graph, cross-hub citation rollup.
  • Auto-synthesis with diff-and-accept UI, confidence tags, hub log, hub lint.
  • Time-traveling hub: /hub/<slug>?at=<date> shows past state.
  • Shared bundles with discoverable opt-in.
  • Permission-aware AI fetching: restricted resources return structured markdown errors, not 404 walls.
  • MWBench cross-AI verification: 100% accuracy across Claude, OpenAI, and Gemini, including on truly unseen content. See /mwbench.

Queued

  • Bundle metadata auto-re-embed on PATCH (currently manual).
  • Cross-encoder reranker on top of RRF for higher chunk precision (when hub sizes warrant the extra 50 to 100 ms latency).
  • Public flip from / (editor) to two-door landing. Held for end of August 2026.
  • Stripe Pro. Beta is free; pricing decided after launch metrics.

Considered but deliberately deferred

  • Multi-vector / late interaction (ColBERT-style) retrieval. Overkill until hubs reach thousands of docs.
  • Mobile native app. Web works on mobile; native is a distraction at this stage.
  • Workspace / team accounts. Current sharing primitives (allowed_emails per doc and bundle) cover most multi-user needs without team complexity.

Not in scope

  • Becoming an LLM provider, or a chat UI, or a vector DB SaaS. Memory.Wiki publishes; it doesn't host inference or own the storage layer.
  • Building a vendor-locked integration. Every Memory.Wiki URL is meant to be readable by any AI; that contract trumps any one-vendor optimization.