Engram 🧠

Long-term memory for your AI. A private, git-backed knowledge base your Claude (or any MCP client) reads and writes across every session and device — so a new chat picks up exactly where the last one left off. Multi-user: invite people, share scoped access to parts of your brain, and let their AI read it directly.

Engram is data + coordination, not an AI. It stores markdown, does git, computes local embeddings for search, and coordinates people. All the intelligence runs on your own Claude/ChatGPT through the MCP tools — nothing is sent to a server-side model. Your memory is a git repo you own, portable across model vendors.

---

What it does

Single-user (the core):

• Remember across sessions — decisions, specs, notes, people, written as work happens (kb_write); load a project's full state in any new session (kb_load).
• Navigate, never ingest — orient from indexes, fetch one concept at a time (kb_read), hybrid semantic + text search (kb_search, local fastembed + optional Qdrant).
• Pass notes between sessions — leave a message the next session must act on; keep a shared running todo list; each project's context.md tracks open loops automatically.
• Artifacts — documents built from knowledge with provenance + staleness tracking + revocable public share links.
• A web explorer — browse projects, concepts, an interactive concept graph, activity, and a live multi-session "office."

Multi-user (v2, opt-in behind ENGRAM_MULTIUSER):

• Open signup — anyone can create their own Engram at /join: OAuth only (GitHub/Google, no passwords), pick a @handle, get a private brain. Operators can close signups (ENGRAM_OPEN_SIGNUP=0) and invite by email or GitHub username instead.
• Publish what you choose — everything is private by default; visibility: private | contacts | public per concept, with a project-level default. Every surface marks the state (🌐 / 👥 / 🔒) so exposure is never inferred from a missing badge.
• A unified home — every user browses their own brain (projects, concepts, graph, search, activity) with the same rich UI, plus profile, contacts, and notifications.
• Contacts + DMs — mutual-consent contacts gate DMs; messages delivered through each person's own Claude; notifications via email + a Chrome extension.
• Context sharing — grant a contact scoped, boundary-safe read access to part of your brain (kb_share_context); their Claude reads your shelf directly (kb_guest_read).
• Tenant isolation — per-user git brains, mandatory search scoping, an adversarial isolation test suite, per-tenant quotas + rate limits, nightly off-site mirrors.

Team memory (v3):

• Search the team before solving — kb_explore(query) is semantic across everyone's public work: your Claude finds the teammate who already hit the problem, with their actual decision. kb_common_ground(@x) shows the concrete concepts two people share — explainable overlap, never an affinity score.
• Rooms — live joins across brains — open a room with teammates (kb_room_open); your Claudes converse via server-side long-poll (free while idle) AND search each other's room-granted work mid-conversation (kb_room_search / kb_room_fetch — path-scoped, auto-revoked on close, every access an audit turn in the transcript). Rooms carry a goal + turn budget so agent conversations terminate. Closing a room offers its outcome back for a human to accept into their brain — never auto-written.
• Ambient team presence — derived from tool calls, zero setup: who's working in which project (never content), in the widget, the dashboard, and the extension; one-click invisible mode.
• One app, one door — a single MCP widget (Home / Browse / People / Rooms / Office) replaces five cards; the web dashboard mirrors the same five-tab IA; one OAuth account spans the connector, the browser, and the Chrome e