
What is a company brain for AI (and what it really ingests today)
A company brain is the memory your AI agents query, with citations you can verify. What it is, which sources it ingests today, who it suits and what it is not.

Letta is an open-source runtime for building stateful agents; MemoryFirst is cited memory for the agents you already use. Pricing and facts as of 2026-11-14.
Search for "memory for AI agents" and both Letta and MemoryFirst show up, yet they solve different problems. Letta is a platform for building stateful agents: each agent owns its memory and edits it itself. MemoryFirst is company memory with citations, which the agents you already use (Claude, ChatGPT, Cursor) connect to over MCP. This article compares them using data from their public websites, accessed on 2026-11-14, and says where each one wins.
One caveat up front: Letta's current website and repository centre on Letta Code and a git-backed memory filesystem, and its README says the older V1 API server now lives on an archive branch. Our knowledge is limited to what its public pages say; anything we could not check is marked as such.

Letta presents itself as an AI research lab in San Francisco, founded by the creators of MemGPT at UC Berkeley's Sky Computing Lab. Its docs call it "the platform for building stateful agents" and offer two paths: managed deployment on Letta Cloud, or running the complete runtime yourself with the Letta App Server. The main repository (letta-ai/letta) is licensed Apache-2.0.
According to its documentation, an agent's memory lives in MemFS, "a git-backed memory filesystem that they can inspect and edit." There is also "dreaming": background subagents that review recent conversations and update memory without interrupting your work.
MemoryFirst ingests your business's events and documents (Gmail and IMAP, Calendar, Drive, Dropbox, OneDrive, Slack, uploaded files and an events API), indexes them with dense search, BM25 and a graph, and answers with citations to the exact source: page, minute, message. It is served through a remote MCP server (15 tools), a REST API and a stdio server. It is not an agent runtime: memory and retrieval only.
| Aspect | Letta | MemoryFirst |
|---|---|---|
| What it is | Platform for building stateful agents | Company memory consumed by other people's agents |
| Who uses it | Developers building agents | Teams and founders using existing agents |
| Open source | Yes, Apache-2.0 (letta-ai/letta repository) | We do not claim it: there is no self-hosted edition of the product |
| Self-hosting | Yes (Letta App Server, letta server) | No: managed service only |
| Memory | Per agent, in git-backed MemFS; background "dreaming" | Shared across the organization, temporal graph with contradictions resolved |
| Citations to the source | We could not verify this on its pages | Yes, span_ref per source type |
| MCP | Appears in its docs; scope not verified by us | Remote MCP server (OAuth 2.1 or API key) and stdio |
| Data connectors | Not verified; its "channels" (Slack, Telegram, Discord, WhatsApp, Signal) are for talking to the agent | Gmail, IMAP, Calendar, Drive, Dropbox, OneDrive, Slack, file upload |
| Data location | Not stated on its pricing page | Database, embeddings and files in the EU; default inference is not EU-only |
| Bring your own LLM key | Yes, on every plan | Yes, for queries and state regeneration |
Letta's figures come from its pricing page (accessed 2026-11-14):
| Plan | Price | What it includes |
|---|---|---|
| Free | $0/month | Up to 3 stateful agents, 5 MiB per memory repository, your own API keys |
| Pro | $20/month | $20 in credits, up to 20 agents, 50 MiB per memory repository |
| Teams Pro | $20/seat/month | Everything in Pro, plus teammates, shared agents and access control |
| Enterprise | Custom | Volume pricing, RBAC, SAML/OIDC SSO, dedicated support |
The page itself warns that real usage depends on the model: chat users often stay within the included credits, casual coding use runs about $100 a month, and power users often exceed $200 a month. With your own keys, spend goes to your model provider.
MemoryFirst's figures come from our pricing page: Free (1 project, memory over MCP and Claude Code hooks, 30-day retention), Pro at 12 EUR a month billed annually (144 EUR a year; monthly billing costs 20% more) with unlimited projects and retention, all connectors, a temporal graph with citations and the morning brief, and Team at 39 EUR a month billed annually (468 EUR a year) with 5 seats and shared team memory. VAT included.
The numbers are not equivalent: Letta charges by agents and model credits, we charge a flat plan. We could not verify total cost on Letta Cloud versus self-hosted Letta.
Pros: open source under a permissive licence; can be self-hosted; full control of the runtime; memory the agent edits and reorganizes on its own; BYOK on every plan; strong research pedigree.
Cons: you must build and maintain the agent; in the texts we read there is no verifiable-citation feature that points back to the source; the pricing page states neither data location nor certifications; cost with hosted models can grow with usage.
Pros: no SDK to maintain; citations to the exact source; memory shared across the team; works with any MCP client; data in the EU; low fixed price.
Cons: no self-hosting; by default the query text and retrieved passages go to Cerebras (US) to write the answer; it is not an agent runtime and cannot be used to build one; some sources (Linear, Sentry, audio) are a plan, not a fact; tenants are separated by query filters plus an optional second barrier, not by separate databases.

For wider context, read the AI memory market in 2026 and how to choose a memory layer, which covers Mem0, Supermemory and Zep. You can also read the comparison with Glean, our approach for coding agents and the plans.

All accessed on 2026-11-14. Anything not listed here could not be verified.
The Free plan includes one project, memory over MCP and the Claude Code hooks.
Start free
A company brain is the memory your AI agents query, with citations you can verify. What it is, which sources it ingests today, who it suits and what it is not.

What it costs when an AI agent starts each session with no past: minutes, tokens and reopened decisions, with an illustrative sum and the market options.

Mem0, Supermemory, Zep, MemoryFirst. They all promise 'memory for AI'. The differences that actually matter: where your data lives, whether you can audit answers, and which model reads your text.