Get started

Engram is your organization's governed memory layer: it captures AI knowledge from daily work, routes it through human approval, and serves it back — permission-trimmed and cited — to employees, applications, and AI agents. This page takes you from zero to a working memory in about ten minutes.

Quickstart#

1. Sign in and enter your workspace

   Go to /app and sign in. What happens next depends on where Engram runs:

   • Engram Cloud — you create (or pick) your organization; its private workspace is provisioned instantly, and organization admins run it.
   • Self-hosted — the deployment has one workspace, and the first user to sign in becomes its owner (see Self-hosting).

   Either way you land in the control plane — a canvas view of your organization as a tree of units, teams, people, and memory shelves.

2. Let the assistant import your knowledge

   A fresh workspace greets you with a short tour that hands off to chat. Attach your documents there — files, a zip, even a spreadsheet where each row is a policy — and the assistant reads them, proposes folders and classifications, and imports once you confirm. Nothing is written without your approval, and placement follows your explicit mapping, never a model's guess.

3. Shape the org tree

   Every workspace starts with a root and starter folders (for example /company/policies, /departments, /projects) plus whatever the import created. Add the units and teams that mirror your company. Folders are also permission boundaries — the tree you draw here is the access model you get.

4. Invite your team

   On Engram Cloud, invite by email (top bar → New → Invite member) and pick a role; invitees land in your workspace only. On a self-hosted deployment, share the app URL — anyone who signs in joins as a member. Members join with a baseline grant: contributor at the root, so a flat team gets a working system with zero configuration. Tighten later with per-folder grants and restricted folders (see Admin & governance).

5. Capture a memory, approve it, ask about it

   Paste a prompt that worked, a decision with its reasoning, or a workflow — from the control plane, the ⌘K omnibar, or the chat panel. The Gardener (Engram's compiler agent) classifies it and files it as a draft — AI-written content always enters as a draft, no exceptions. Approve it in the review queue, then ask the chat panel a question it answers. You get a cited answer — every claim names the memory it came from — or an explicit refusal if nothing accessible covers it. No free-association, ever.

6. Connect a tool

   In Settings → Integrations, connect Linear, GitHub, Jira, Slack, Notion, or Google Drive. Synced content lands as drafts (or pre-approved, if an admin marks the corpus trusted) in the folder you choose. Details in Connectors.

7. Connect Claude to your memory

   Mint an API token in Settings, then from any machine running Claude Code:

   claude mcp add engram --transport http \
     https://<your-deployment>/api/mcp \
     --header "Authorization: Bearer egm_..."

   Your Claude session can now ask_memory, recommend_for_task, and capture_memory against your workspace — fenced by the exact same permissions as every other surface. Full setup in MCP.

What you just built#

In ten minutes you exercised the whole loop that makes Engram different from a wiki or a chatbot:

• Capture is dumb and fast; compilation is smart and async. You threw raw text at the system; an agent did the filing.
• AI writes are always drafts. The Gardener contributed to memory, but only you — a human with manage rights on the folder — promoted it into the trusted corpus.
• Permissions live in the query. The answer you got was trimmed to your access in SQL, not filtered afterwards. Ask about something you can't see and you get a refusal, not a leak.
• Agents are ordinary callers. The Claude session you connected has exactly your token's power — no privileged path, and prompt injection dies at the SQL fence.

Where to go next#

• Core concepts — the five ideas everything else builds on: memories, the tree, the lifecycle, the access algebra, cited answers.
• The control plane — a tour of every surface: canvas, chat, omnibar, review queue, analytics.
• Connectors — data in: org-level and personal connections for six providers plus bulk upload.
• MCP — memory out: all 33 tools, per-client setup, and the permission story.
• API reference — the /v1 surface, on-behalf-of identity, webhooks, and the TypeScript SDK.