Linked Layer

Shared memory for teams & agents. A token-gated context layer over all your tools — collected into a permission-aware graph and served to people and AI agents in a single call: recall(query, scope).

🌐 linkedlayer.xyz · ⛓ Solana

---

The problem

A team's knowledge is scattered across Slack, GitHub, Notion, Drive, Linear and call transcripts. The why behind decisions lives in someone's head or buried in a thread. New hires spend weeks reconstructing context, decisions get silently re-litigated, and AI agents act on stale or hallucinated information.

Linked Layer turns that scattered activity into one living, permission-aware memory that both people and agents can query.

How it works

1. Connect sources — Slack, GitHub, Notion, Drive, Linear & more ingest into one place; permissions mirrored from each source.
2. Build the graph — a permission-aware context graph of projects, people, decisions and threads, kept current by incremental sync.
3. Distill — an LLM continuously extracts decisions, the "why", action items and statuses (deduped).
4. Recall — people ask in plain language; agents call recall() over MCP. Same memory, same permission bounds.

Key features

• Permission-aware by default — retrieval is filtered through each item's source ACL at query time and fails closed. Nothing is surfaced that the caller couldn't already see.
• One primitive, two audiences — humans ask in a chat; agents call recall() over MCP / the Context API.
• Cited & traceable — every answer links back to the exact source nodes it used.
• Always-current — incremental, deduped sync keeps the graph fresh.
• Token-gated + pay-per-call — hold $LINKED to use the layer; external agents pay per recall() via x402. Fees fuel buyback & burn.

Tech stack

TypeScript · pnpm monorepo · Fastify · Drizzle ORM · Postgres + pgvector · BullMQ · Solana Web3.js · React · Vite · Tailwind · Framer Motion

apps/
  web/         landing + "ask the company" chat (Vite + React + TS)
packages/
  core/        domain types, graph model, zod schemas, config
  db/          Postgres + pgvector (Drizzle), hybrid search
  embed/       embeddings provider abstraction (Voyage | stub)
  connectors/  GitHub, Notion, Slack + connector interface
  distill/     LLM distillation — decisions / why / action items
  gating/      Solana SPL token gate + Sign-In-with-Solana + x402
  engine/      orchestration: ingest → distill → embed → recall
  api/         Fastify Context API + OpenAPI/Swagger
  mcp/         MCP server — recall / search / write
  worker/      BullMQ background workers + scheduler

Quickstart (local dev)

pnpm install
cp .env.example .env          # LLM/embedding keys are optional

docker compose up -d          # Postgres + pgvector + Redis
pnpm db:migrate

pnpm dev                      # API + worker
pnpm web                      # frontend on :5173
pnpm test                     # vitest

No LLM key? A heuristic fallback keeps the pipeline running. No embedding key? Stub embeddings work out of the box — zero hard dependencies for local development.

MCP — plug into any AI agent

{
  "mcpServers": {
    "linked": {
      "command": "npx",
      "args": ["-y", "linked-layer-mcp"],
      "env": { "RECALL_API_KEY": "your-key" }
    }
  }
}

Your agent now has recall(), search() and write() — grounded in your team's real history, bounded by its real permissions.

Roadmap

• Discord & Telegram connector
• Multi-workspace support (cross-org recall with ACL firewall)
• Streaming recall via SSE
• $LINKED staking tiers — higher rate limits & priority indexing
• On-chain proof of recall (Solana attestation per answer)
• Self-hosted deployment guide (Kubernetes helm chart)

License

MIT