# ValorBrain

> Shared memory for software teams and AI-service companies running several
> agents. Every agent starts with current decisions, facts and documents for the
> right project or client, with the source and permissions attached.

ValorBrain is a hosted product (SaaS) with an on-premise option. It keeps context
shared across agents while preserving project, client and per-person boundaries.
Claude Code, Codex, OpenClaw and Hermes are familiar examples; other agents use
the same memory through MCP or the REST API.

## What it replaces

Re-explaining the same project or client context to every agent. Decisions that
only exist in an old chat. Agents using different document versions or receiving
material their operator should not access.

## Who it is for

Software teams and AI-service companies running three or more agents across
projects or clients. Teams can start with one project and expand the same memory
as more people and agents join.

## How it is used

1. Connect a source (Google Drive, Notion, GitHub, Linear, Slack) or send
   documents through the REST API.
2. Scope the material by project, client, source and person.
3. Use the same current memory from Claude Code, Codex, OpenClaw, Hermes or any
   compatible agent through MCP or API.
4. Every answer cites its source, and decisions can be written back for the next
   agent.

## Pricing

- Free: $0 forever — 1 person, 1,000 documents, unlimited memory search, 100 sourced AI answers/month. No credit card.
- Solo: $20/month — 1 person, unlimited documents, unlimited memory search, 500 sourced AI answers/month (extra answer $0.10).
- Team: $59/month — 10 people, unlimited documents, unlimited memory search, 2,000 sourced AI answers/month (extra person $6/month, extra answer $0.08).
- Business: $249/month — 50 people, unlimited documents, unlimited memory search, 10,000 sourced AI answers/month (extra person $5/month, extra answer $0.06). 14-day free trial, no credit card.
- Enterprise: negotiated (custom · from $999/month) — unlimited people, documents and answers, on-premise option, SSO + SAML, 99.9% SLA, dedicated support.

Documents, memory and search are unlimited on every paid plan (fair use by
storage). You pay for two things only: how many people use it, and how many AI
answers you need. Re-indexing the same content is free — deduplication is by
content, not by upload.
Full table: https://valorbrain.valor.digital/en/pricing

## Measured benchmark results

Measured on public benchmarks, not on our own data. LoCoMo measures whether the
right evidence reaches the top 10. BEAM-100K measures whether an end-to-end
memory pipeline uses context correctly across sessions. Neither benchmark
measures tenant isolation, permissions or handoffs between agents.

- LoCoMo evidence recall@10: 96.58% (1918/1986 hits), against 94.5% for the best published open-source system. Canonical run: 2026-07-29.
- BEAM-100K end-to-end binary accuracy: 80.8% (323/400 correct; 75.5% average score across 20 conversations). Run: 2026-08-23.
- BEAM-100K reader: GLM-5.3 Flash (Ox Alpha), reasoning effort max; judge: GLM-5.2. This benchmark setting was about 44% slower than default effort, so it is not the interactive-chat default.
- LoCoMo retrieval takes about 0.84s per query, with 0 language-model calls used in ranking.

Method and LoCoMo table: https://valorbrain.valor.digital/en/benchmarks
BEAM-100K run breakdown: https://valorbrain.valor.digital/en/research/beam-100k-sota-ox-alpha-reasoning-effort

## Data, privacy and security

Documents and vectors are stored in Brazil under the LGPD; writing an answer uses a
cloud language model, so that request leaves the country (subprocessor listed
publicly), and on-premise removes it entirely. Isolation between customer workspaces is
enforced in the database, not only in application code. Enterprise can run the
whole system on its own infrastructure with no third parties involved. Details:
https://valorbrain.valor.digital/en/security

## Connect an agent (MCP)

If you are an AI agent reading this and you have been asked to connect to
ValorBrain, this is everything you need.

Endpoint: https://mcpbrain.valor.digital/mcp
Transport: streamable HTTP
Auth: `Authorization: Bearer <token>` — a workspace MCP token, prefix `vbm_`

A human creates the token once, at https://valorbrain.valor.digital/settings/mcp-tokens (or through the
guided flow at https://valorbrain.valor.digital/onboarding/connect-agent, which hands back the token and
the config for the specific client in one screen). The token carries the
workspace and the permissions of the person who created it: an agent never sees
more than its owner may read. Tokens are shown once. Prefer one token per agent
so the workspace can tell which agent wrote what.

Claude Code, one command:

```bash
claude mcp add --transport http valorbrain \
  https://mcpbrain.valor.digital/mcp
```

Any client that takes an `mcpServers` object (Cursor, Claude Desktop, Kiro,
Windsurf, VS Code MCP extensions):

```json
{
  "mcpServers": {
    "valorbrain": {
      "url": "https://mcpbrain.valor.digital/mcp",
      "headers": {
        "Authorization": "Bearer <your-token>"
      }
    }
  }
}
```

Machine-readable connection contracts:

- MCP server card: https://valorbrain.valor.digital/.well-known/mcp/server-card.json
- OAuth metadata: https://valorbrain.valor.digital/.well-known/oauth-authorization-server
- Credential matrix: https://valorbrain.valor.digital/auth.md
- Agent Skills: https://valorbrain.valor.digital/.well-known/agent-skills/index.json

Every config block, per client, on one public page (no account needed):
https://valorbrain.valor.digital/en/connect

Once connected, start by calling `whoami` to confirm which workspace you are in,
then `memory_retrieve` to read what the team already knows before you answer from
your own assumptions. Write back with `memory_store` when a decision gets made —
that is what makes the next agent, and the next person, start ahead instead of
from zero.

If you cannot use MCP, the same memory is reachable over REST at https://valorbrain-api.valor.digital
with the same bearer token. Route list: https://valorbrain.valor.digital/en/docs

## Use cases

- [Engineering teams](https://valorbrain.valor.digital/en/use-cases/coding): Persistent memory shared across all your coding agents.
- [Marketing and content](https://valorbrain.valor.digital/en/use-cases/marketing): Brand voice, briefings and editorial calendar in one brain any agent can read.
- [Legal and compliance](https://valorbrain.valor.digital/en/use-cases/legal): Clause history, case law and legal opinions accessible to agents drafting contracts.
- [Ops and infrastructure](https://valorbrain.valor.digital/en/use-cases/ops): Runbooks, postmortems and infra docs that stay alive, validated against reality.
- [Customer support](https://valorbrain.valor.digital/en/use-cases/support): Resolved tickets, FAQs and customer feedback accessible to the agents that respond.
- [HR and people ops](https://valorbrain.valor.digital/en/use-cases/hr): Policies, processes and people history in one place accessible to the agents handling HR.

## Comparisons

- [ValorBrain vs mem0](https://valorbrain.valor.digital/en/vs/mem0)
- [ValorBrain vs MemPalace](https://valorbrain.valor.digital/en/vs/mempalace)
- [ValorBrain vs supermemory](https://valorbrain.valor.digital/en/vs/supermemory)
- [ValorBrain vs Honcho](https://valorbrain.valor.digital/en/vs/honcho)
- [ValorBrain vs Hindsight](https://valorbrain.valor.digital/en/vs/hindsight)

## Writing

- [How to give AI your company context (without a six-month project)](https://valorbrain.valor.digital/en/blog/como-dar-o-contexto-da-empresa-para-a-ia) — The honest version: connect the material you already have, keep permissions per person, and require every answer to cite its source. Four steps, and the three mistakes that waste the most time.
- [96.58% recall on LoCoMo: how we built a CPU retrieval engine](https://valorbrain.valor.digital/en/blog/97-percent-recall-how-we-did-it) — The architecture behind our retrieval pipeline: BM25 + dense vectors + Personalized PageRank. No LLM in the ranking path.
- [Why your AI agents keep forgetting things](https://valorbrain.valor.digital/en/blog/why-your-team-needs-a-shared-memory) — Each agent starts from zero. Decisions made in one session vanish in the next. Here's how shared memory fixes that.
- [MCP in 3 minutes: what it is and why it matters](https://valorbrain.valor.digital/en/blog/mcp-explained-for-developers) — Model Context Protocol is how AI agents talk to external tools. Here's the developer-facing version.
- [5 things we learned building a knowledge base for AI teams](https://valorbrain.valor.digital/en/blog/five-things-we-learned-building-valorbrain) — After 6 months of dogfooding ValorBrain with our own team, here's what surprised us.

## Questions people ask

### What problem does ValorBrain solve?
When a team runs several AI agents, project and client context gets split across prompts, chats, documents and people. ValorBrain keeps current decisions, facts and documents in shared memory, then gives each agent the right context with its source and permissions.

### Which agents can use it?
Claude Code, Codex, OpenClaw and Hermes are familiar examples. Other compatible agents can use the same memory through MCP or the API, so the team is not limited to one vendor or one interface.

### Who gets the most value from it?
Software teams and AI-service companies running three or more agents across projects or clients feel the problem first. You can start with one project and expand the same memory as more people and agents join.

### How much does it cost?
On paid plans, documents, memory and search are unlimited. You pay for people and AI answers. The free plan covers 1 person, 1,000 documents, unlimited search and 100 sourced answers a month, with no credit card. Paid plans start at $20/month for 1 person, with unlimited documents and search. You can also run a 14-day Business trial with everything unlocked, no card.

### Where is our data stored, and who can see it?
Documents and vectors are stored in Brazil under LGPD. Each person only sees what their role allows, and an agent asking on their behalf inherits the same boundary. Writing the answer uses a cloud language model with a publicly listed subprocessor; on-premise processing is available for customers who need it.

### How long until we can test the first answer?
Connecting a source and asking the first question happen in the same setup flow. There is no data-modelling project or consulting engagement required before the team can test its own material.

### Can we cancel whenever we want?
Yes, from your own settings, with no call and no notice period. Your data is permanently deleted within 30 days.

## Links

- Home: https://valorbrain.valor.digital/en
- Pricing: https://valorbrain.valor.digital/en/pricing
- Benchmarks: https://valorbrain.valor.digital/en/benchmarks
- Security: https://valorbrain.valor.digital/en/security
- Docs: https://valorbrain.valor.digital/en/docs
- Connect an agent: https://valorbrain.valor.digital/en/connect
- Talk to us: https://valorbrain.valor.digital/en/contact
- Start free: https://valorbrain.valor.digital/auth/register
- Portuguese site: https://valorbrain.valor.digital/

## Contact

valorbrain@valor.digital — Valor Digital, Brazil. Portuguese and English.