Project memory concepts and developer references.

What Shelvia is, and what it isn't.

Bring in your first work and approve your first memory.

Reviewed decisions, sources, prompts, next steps, and constraints inside one workspace.

AI-generated, imported, and agent-proposed memory enters review before it becomes trusted. Human-approved project edits are audit logged.

Ranked context for the next session. Source-backed and health-aware.

Stale, conflicted, low-confidence, and missing-source memory surfaced with explicit verdicts.

GitHub, Notion, and Google Drive / Docs feed the review queue with cost guardrails.

Hybrid keyword + vector search across approved memory, workspace-scoped at every layer.

Deterministic compression of older approved memory into durable project state.

User-initiated capture from ChatGPT, Claude, Perplexity, Gemini, and NotebookLM, routed through review.

Recipes, candidate-write APIs, MCP tool exposure, SDK helper, and observability logging. Agent, API, and MCP writes enter review before becoming trusted memory.

Workspace tokens and OAuth 2.0 with PKCE.

Read memory, generate context packs, create review candidates, log continuations.

Model Context Protocol endpoint that honors the same review-before-save contract as REST.

Typed client wrapping projects, search, candidates, context packs, summaries, and webhooks.

Signed outbound delivery on candidate creation, context-pack generation, and handoffs.

The integration surface is reachable now. This roadmap entry tracks the manual four-runtime attestation that gates full "live" status.

Notable shipped changes, in plain language.