Company Brain

"A teammate in your Slack that truly knows and understands your company, and can do anything." An open-source Slack agent from supermemory (supermemoryai), TypeScript, Apache-2.0. Read 2026-09-28.

The README's own framing is the notable fact before any feature: this used to be a paid product with thousands of users, was discontinued "a few weeks ago," and is now released in full, reworked to deploy on the user's own Cloudflare account rather than supermemory's. A shipped commercial Slack bot going fully open, not a from-scratch open-source release.

What it does

It sits in Slack channels, absorbs what the team discusses — decisions, projects, owners, context — so nothing has to be written down separately, then answers questions grounded in that memory plus whatever connected tools cover, and can act: opening a GitHub issue or a Linear ticket itself rather than just describing what to do. It also joins conversations unprompted when it has something relevant to add (chattiness is configurable per-org and per-channel), runs scheduled digests, and can start research on its own. A worked example in the README: asked "is prod down?", it answers unprompted, citing 42 errors in the last 15 minutes.

Permission model — a scoped memory graph

Memory is modelled as a permissions graph, not one shared pool:

Public channels draw on the shared org-wide brain.
Private channels add that channel's own memory on top.
DMs add the asker's personal memory plus every private channel they belong to.

The write side carries the same scoping one step further: a tool write runs under the asking user's own connected-tool credentials, not the bot's. If a request needs a teammate's tool instead, the bot sends that teammate an approve/deny card first — "nothing is granted silently." That is a concrete instance of agent-permission-model (approval on the specific action that needs someone else's grant, not a blanket per-step prompt) and a second data point for agent-scope-isolation after qm — another multiplayer Slack agent, another per-channel/DM memory partition, another "reads only with the asking user's own access" rule. Where qm states the containment position explicitly ("the agent is not trusted to make authorization decisions," enforced by a separate core) and lists its own known gaps in a 9.9 KB SECURITY.md, Company Brain's README states the same rule as a feature claim with no comparable limitations section — read the scoping as a design commitment demonstrated by neither repo's usage in production yet, not a verified guarantee.

Tools and runtime

MCP integrations — GitHub, Linear, Notion, Google Workspace, and, per the README, "hundreds more" over MCP.
Sandbox — clones repos, runs scripts, returns charts, CSVs and PDFs into Slack. QuickJS runs a "Code Mode" inside the Worker itself; the heavier sandbox is a Cloudflare container (Workers Paid) or Daytona (free plan).
Skills — teachable processes, formats and house voice, plus one workspace-wide prompt for general behavior. Undocumented beyond naming; no worked example of authoring one.
Web UI — a live graph of what the bot remembers, alongside settings for tools, models, proactivity, automations and skills.

Infrastructure and memory backend

TypeScript on Cloudflare Workers, Durable Objects, D1, KV and Workers AI; Drizzle for schema and migrations; Bun for tooling. Memory itself is stored on supermemory, the vendor's own hosted service — so the application is self-hosted (one-click deploy to the user's Cloudflare account) while the memory substrate stays a hosted third-party dependency, a split agent-memory hadn't seen yet: the hosted/self-hosted axis there (cloudflare-agent-memory vs. everything self-run) assumed the two travel together. Deploy needs exactly two secrets: SUPERMEMORY_API_KEY and a MODEL_API_KEY (Anthropic, OpenAI, Google, xAI or OpenRouter — provider auto-detected, billed directly, no markup). The free Cloudflare plan caps each request at 50 outbound calls, which can truncate long answers; Workers Paid ($5/mo) is the README's recommendation, and on the free plan the sandbox additionally needs a DAYTONA_API_KEY.

Cross-spoke context — the PKM angle

The hub's router also weighed ../research-wiki, where the memory/team-knowledge framing has a close relative: gbrain (cross-wiki), Garry Tan's open-source "brain layer," runs a self-wiring knowledge graph over MCP as a personal/team markdown knowledge base rather than a Slack agent's grounding memory. Company Brain is the packaged-agentic-product reading of a similar "team memory" idea — the dominant substance here (Slack bot, tool orchestration, MCP actions) is agentic-tooling's domain per the 2026-06-01 split, so it's paged here with the knowledge-management angle noted rather than split across both spokes.

Reading the evidence

717 stars, 105 forks, 42 commits, 2 open issues, 9 pull requests at read time — a young repo by this corpus's standards, and every claim above (the permission model, the unprompted-reply example, "five minute" deploy) is the maker's own README with no independent usage report or security audit found. No stated publication date for the repo itself (the README says the product was discontinued "a few weeks ago"; the page footer shows © 2026) — left off rather than guessed.

agent-permission-model · agent-scope-isolation · agent-memory · qm · cloudflare-agent-memory · model-context-protocol · supermemory · gbrain · agentic-tooling · synthesis