Company Brain for people and AI agents
Context shouldn't start over with every agent.
ValorBrain is the shared memory that keeps work moving when the person, the agent, or the model changes. Decisions, rules, and handoffs get captured in the flow and handed to whoever comes next, with source, permission, version, and history.
- LoCoMo · R@10
- 96,58%
- BEAM-100K · R@10
- 80,8%
- MCP tools
- 90+
Internal retrieval benchmarks; not peer-reviewed, not end-to-end QA. See caveats
Data in Brazil · LGPD · tenant isolation
Context unit
- id
- CTX-2026-ENG-007
- version
- v1 · current
- source
- ADR-007 · PR #412
- permission
- eng/*
01The problem
Your team already explained this once
With three or more agents in operation, context fragments faster than any wiki can keep up. The cost never shows up as a budget line: it shows up as repeated briefings, outdated answers, and rework.
- 01
Fragmented context
The decision lives in the Claude Code chat, the why lives in the PR, and the exception lives in the head of whoever was in the meeting. Cursor sees none of it.
“Where was that decided, again?”
- 02
Repeated briefings
Every new agent asks for the same context. The team pastes the same PDF, the same ADR, the same policy. Again. For the third agent this week.
“Let me paste the doc again.”
- 03
Stale documents posing as answers
The rule changed on Tuesday; the agent answers with the March version. Nobody notices until the client, the auditor, or the person on call does.
“Didn't this rule change?”
- 04
Permissions rebuilt from scratch
Every tool has its own access control. Adding one more agent turns into a governance project that starts from zero.
“Is this agent allowed to see finance?”
Same team, two context architectures
ADR-007 buried in chat history
test convention pasted by hand
has no idea ADR-007 exists
briefing rebuilt from zero
March's INFRA.md
no access to team decisions
re-explains everything in standup
reviews permissions tool by tool
Four partial views of the same work. No current version anywhere.
02How it works
Context is born from the work. And returns to it.
No modeling project before you start. ValorBrain captures what already happens in agent sessions and team tools, structures it into living memory, and hands it to whoever needs it next, with governance applied before the prompt.
01 · Capture
Capture without friction
Hooks in the agents' lifecycle record decisions, open items, and handoffs while the work happens. Documents, transcripts, and the REST API complete the picture. Context comes from the work itself, not from a modeling project.
- Automatic hooks in agent sessions
- Document and transcript ingestion
- REST API for internal systems
- BYOK/OAuth connectors as sources: GitHub, Notion, Slack, Google Drive, Linear, HubSpot, Jira, Granatum
Automatic hooks
- context-surfacing
- decision-extractor
- handoff-generator
- feedback-loop
- precompact-extract
- postcompact-inject
- curator-nudge
// Claude Code session · eng-corehook: decision-extractoragent: claude-codesession: eng-core/2026-09-22→ decision captured id: CTX-2026-ENG-007 title: "Cursor pagination for public APIs" source: "ADR-007.md · PR #412" author: ana (tech lead) status: awaiting human validationcontext-surfacing
Brings relevant context in before the agent answers.
decision-extractor
Records decisions made in the session, with source and author.
handoff-generator
Generates the handoff when work moves to another agent or person.
feedback-loop
Turns human corrections into lessons for future sessions.
03Multi-agent coherence
Context continues when the agent changes.
Claude Code captures, Cursor sees it on the next query. Hermes ingests overnight, and the morning agent gets a bootstrap of what changed. Handoffs between agents are routed, not lost in chats.
Claude Code captures. Cursor sees it on the next query.
Illustrative example- Tue 16:42Ana · tech lead
“Let's standardize cursor pagination for public APIs. Offset is superseded.”
- Tue 16:42Claude Codedecision-extractor
Decision recorded as CTX-2026-ENG-007 · v1, with source in ADR-007.md and PR #412.
- Wed 09:12João · dev
“Creates the invoice listing endpoint.”
- Wed 09:12Cursorcontext-surfacing
Endpoint created with cursor pagination, per ADR-007.
[Fonte: CTX-2026-ENG-007 // v1]
Nobody re-explained the pattern. Cursor started from the previous day's decision.
Lives where the work already happens.
Hooks and MCP in the agents' cycle; REST API for everything else. The team stays in the tools it already uses. ValorBrain doesn't ask for one more screen.
Agents · via MCP and API
- Claude Code
- Cursor
- Kiro CLI
- Codex
- Devin
- Hermes
- OpenClaw
- ChatGPT
- Gemini
- Internal agents via MCP/API
Sources · BYOK/OAuth
- GitHub
- Notion
- Slack
- Google Drive
- Linear
- HubSpot
- Jira
- Granatum
The connector is a means. What reaches the agent is context, with source, permission, and version.
04Governance
Governance before the prompt.
Per-person and per-source permissions applied at the semantic layer. The agent inherits access from whoever uses it; sensitive domains stay isolated; every read is traceable. When context is missing, the answer says so and points to the person responsible.
Simulator · same question, different access
Illustrative exampleQuestion asked to the agent
“What's the maximum discount without executive approval?”
10% without executive approval. Above that, RevOps validation is required.
[Fonte: CTX-2026-GTM-904 // v3]
▸inherit(user=marina)
→scope: gtm/pricing ✓
→CTX-2026-GTM-904 · v3 · current
→read-log recorded
CTX-2026-GTM-904
Discount policy for quarterly closings
- in force since
- 2026-09-15
- approved by
- commercial board
- v2
- superseded, kept in history. The next query already starts from v3.
Inherited permissions
The agent only accesses what the person using it can see, per person and per source.
Tenant isolation
Multi-tenant RLS. Sensitive domains, like tax or M&A, stay isolated.
Versioning and supersession
The new rule supersedes the old one without erasing history. The next version starts from there.
Conflict and contradiction
Sources that disagree get flagged for human review before they become an answer.
SHACL quarantine
Records that violate the schema stay out of answers until they are fixed.
Provenance chain
From the original source to the final answer, every step is logged.
Drift and reality probes
Documentation is compared against the real system; divergence becomes an alert, not a surprise.
Read auditing
Who read, through which agent, at which version. Every read is traceable.
Operations and trust
- Data in Brazil · LGPD
- Tenant isolation
- Permissions per person and source
- Auditing of every read
- No outbound calls from the engine
Connectors live in the SaaS; the engine makes no outbound calls. Human review where the decision is sensitive.
05Verticals
The same core, in your team's flow.
The mechanism is the same; what's worth remembering changes. Each vertical comes with a scenario, typical agents, a schema pack, and the indicators that make sense to track. No invented targets.
Department · pack:
“The customer shouldn't have to repeat themselves. The agent should remember.”
The pain
Exceptions, cases, and escalations with source and inherited permission: the customer doesn't repeat the case and the agent answers by the current version.
Typical agents
- Support agents
- Service copilots
- Internal bots
Scenario
- 01
The support agent asks about the case and receives the decisions and exceptions already granted to that account, like the extended refund deadline on ticket #8812, with source, who approved it, and how long it holds.
- 02
Permissions are inherited from whoever uses the agent: it sees only what that person can see.
- 03
The answer cites the source of the precedent instead of reconstructing the case from memory.
- 04
A new ticket retrieves the resolutions of similar cases, with the outcome of each.
- 05
What the team learned stays on record as a learning, with the originating case cited.
- 06
The agent makes a proposal with precedent in hand, not a guess.
- 07
The case escalates to the supervisor with context ready: history, exceptions in force, and sources.
- 08
The customer doesn't repeat the story to the second person.
- 09
Whatever the supervisor corrects becomes a recorded lesson for the team's and the agents' next sessions.
Schema pack
- customer
- case
- escalation
- policy_exception
- resolution
- learning
What to track
- Resolution time
- Context rework
- Consistency across channels
06Who decides
One layer. Five different questions.
Whoever buys context infrastructure isn't always whoever writes code. The same layer answers the engineer and the COO, each with the proof that matters to their decision.
What they decide
Architecture, isolation, integration, and the cost of re-explaining everything to each agent.
The objection we hear
“We already have a wiki, Notion, and RAG.”
The answer
A wiki is static and RAG only answers. ValorBrain versions, cross-references sources, applies inherited permissions, and delivers over MCP to any agent, with published benchmarks and their caveats.
- Inherited permissions
- MCP · 90+ tools
- Cross-agent
- Benchmarks with caveats
07Proof
Numbers with the caveats next to them, not in a footnote.
If the right context doesn't reach the agent, it starts over. That's why we measure retrieval: high recall is the precondition for continuity, not a guarantee of a correct answer.
LoCoMo
pipeline completo · R@10 · +2,1pp vs líder open-source
96,58%
BEAM-100K
respostas end-to-end · acurácia binária
80,8%
Benchmark configuration: reader GLM-5.3 Flash (Ox Alpha); judge GLM-5.2; effort=max; run 2026-08-23; ~44% slower than default effort; the reader is the customer's choice — results vary with the model (stronger readers can score higher). ValorBrain doesn't sell LLM calls: the memory layer is the product
Read this before citing
- These are internal retrieval benchmarks. Not peer-reviewed.
- They don't measure end-to-end QA: final answer quality also depends on the model and the prompt.
- R@10 tells you whether the relevant chunk appears among the top 10 retrieved results.
- Axes run 0 to 100%, with no truncation.
~90%
of production retrieval comes from hooks
context arrives unprompted
90+
MCP tools in the current catalog
for any compatible agent
How retrieval works
- Question
- Query expansion + intent
- BM25 PT/EN ‖ dense vectors
- RRF
- Personalized PageRank
- Cross-encoder
- MMR
- Context with source
Lexical search in Portuguese and English combined with dense vectors, RRF fusion, personalized PageRank for whoever asks, cross-encoder reranking, and MMR for diversity.
08Pricing & enterprise
Start with one project. Scale when it makes sense.
Prove it in your real workflow before any contract. Once context starts flowing between teams and agents, the scaling conversation starts from what you've already measured.
Free
starting pointOne project
To prove it in your real workflow, with your own agents, before any sales conversation.
- Hook capture in agent sessions
- Living memory with version and source
- Delivery over MCP and REST API
Solo, Núcleo & Operação
from US$20/month
For teams with several projects, areas, and agents sharing the same context.
- Schema packs per area
- BYOK/OAuth connectors as sources
- Handoffs between agents and teams
Enterprise
from R$ 4,997/month
For operations with isolation, auditing, and contract requirements.
- Tenant isolation, permissions per person and source
- Auditing and provenance chain
- Data in Brazil · LGPD
Migrate without rewriting the past.
Documents, decisions, and transcripts that already exist come in through ingestion; versioning starts there. The wiki can keep existing, now as one versioned source among others instead of the team's only memory.
Scope and limits of each plan are confirmed at signup or during the sales conversation.
09Questions
The objections we hear, and how we answer.
No “absolutely!” and no miracle promise. If your question isn't here, it's a good way to start a conversation.
We already have Notion or a wiki.
A wiki is static. ValorBrain versions, cross-references sources, injects context into the agent, and surfaces a conflict when two sources disagree. The wiki can stay, as one source among others.
Isn't this just RAG?
RAG answers a question. ValorBrain keeps decisions, permissions, versions, and handoffs between agents, and knows when an answer went stale because the rule changed.
Will this be one more tool for the team to open?
No. It lives in the hooks and MCP of the agents the team already uses. Capture happens during work; delivery happens at question time.
What about data privacy?
Data in Brazil, in compliance with the LGPD. Tenant isolation, permissions inherited per person and source, read auditing, and no outbound calls from the engine. The connectors live in the SaaS.
What if the agent uses an old version?
Every unit has validity (as_of) and supersession. The current rule arrives before the old one; conflicts and drift get flagged for human review. We don't promise infallibility. We promise traceability.
Do we need to rewrite what we already have?
No. Migration happens without rewriting the past: existing documents, decisions, and transcripts come in through ingestion, and versioning starts from there.
How much work does it take to maintain?
Automatic hooks do the capture; schema packs add structure; memory health and auditing show what has aged. Human curation goes where it makes a difference: validation and review.
Next step
Less starting over. More continuity
Bring a real project. In one conversation we map where context gets lost between the people and agents on your team, and where to start.