The brain your AI was missing. All of your company's know-how in a single brain.
Your AI agent knows how to code, but it doesn't know your company: the why behind every decision is scattered across code, tickets, chat and meetings. Brain brings it together in a database on your infrastructure, with the tools and controls your agent needs to query it. Every fragment comes with a reference to its source.
Watch the video · 1:31Version 0.4.14 · distilled from a production implementation at a 4-year-old fintech.
Connects to every system in your company: GitHub, Jira, Slack, Confluence, Google Drive, meeting recordings, CRM, KYC, e-signature, Google Analytics, Google Ads, Meta Ads, email marketing, push notifications, cloud, Google Workspace, Figma, Clarity and more.
The same thing, seen from IT. In a minute and a half.
The agent that codes well but doesn't know your company, and the same agent with Brain as its brain.
The everyday chat
The question nobody can answer, and the person who knew is gone.
The same situation, with Brain
The answer with its source, in seconds.
The numbers from the real case
What was measured in a production implementation.
Illustrative dialogues · 1:31 · plays muted
this decided?
What it solves
Ten problems that almost every company has.
A key person leaves and takes what they knew
The files stay. The why behind every decision leaves with the person.
Nobody knows why something was decided
The answer is scattered across chats, meetings, documents, tickets, tools and systems.
Everything depends on who's around that day
The same question gets a different answer, or no answer at all.
You pay for things nobody uses
Services and licenses that keep billing because nobody cross-checked the data.
AI makes things up about your company
It confidently answers what it doesn't know, and there's no way to check.
New hires take months to get up to speed
They learn by asking whoever has time to explain.
Manual tasks that keep repeating
People with years of experience stuck on repetitive work that could be automated.
Decisions run on gut feeling
Conclusions and decisions based on assumptions, not hard data.
No time left to improve
The team lives putting out fires: nobody spots what could be automated or knows where to start.
The competition is already moving faster
Whoever works with AI lowers their operating costs and passes it on to their prices.
Brain doesn't replace your team. It replaces depending on someone to remember, and it makes your team stronger.
AI alone
An agent reasons well about what's in front of it. Without a sourced memory, when it comes to your company it fails in eight known ways.
Brain doesn't make AI infallible. It makes its mistakes visible, and keeps corrections from getting lost.
The AI only sees what you paste into the chat: a file, an error, a question.
It can query everything your company has already written, and brings in what each question needs.
What it is
Your company's memory, and the controls that make it auditable. Agents are missing the history: what was asked for, what was discussed and what was decided. Brain gives it to them, with the source of every part.
A database built from your sources, a server that exposes it to your agent, and a set of rules that stop what shouldn't happen. It's not another app: it installs into the AI you already use.
A model of its own
Nothing gets trained. The one doing the reasoning is your AI agent, on your account.
A search engine
It cross-references sources with reasoning, keeps your company's context, and says where it got each part of the answer.
A chatbot over your documents
Beyond answering, it stops what shouldn't happen and leaves a trail for every answer.
A replacement for your team
It makes people stronger: the decisions are still theirs.
not a pile
What a corpus is
The word you'll see most on this site.
From the Latin corpus, “body.” In English its plural is corpora (wiktionary).
- In law
- The Corpus iuris civilis, “body of civil law,” gathered Roman law into a single searchable work under the Byzantine emperor Justinian, between 529 and 534 (es.wikipedia).
- In linguistics
- A collection of real texts, gathered and organized to study how a language is used.
- In artificial intelligence
- The texts a model works on. Often, the ones used to train it.
- In Brain
- Everything your company has already written—tickets, code, chat, meetings, documents—gathered into a single body of evidence. It isn't used to train the model: it gets queried, and every answer says which part it came from.
We say corpus and not “files” because it isn't a pile: it's a body, with parts that touch. The ticket explains the commit, the meeting minutes explain the ticket, and the chat explains why the other option was dropped.
The thesis
AI doesn't need to always say the same thing. You need to be able to verify what it said.
No model is deterministic, and Brain doesn't try to make it one: it changes where the truth it relies on lives, in indexed documents with their citation alongside. What it brings is auditability.
Your AI agent is the developer who shows up knowing how to code but knowing nothing about your system. Brain is what that developer would read in their first six months, if it were written down, in one place, and true.
Tap AI with Brain to see the difference.
Illustrative example with made-up data.
| Auditability | AI without Brain | AI with Brain |
|---|---|---|
| Where a fact comes from | From what the model remembers or what was pasted into that chat | From an indexed document, with its citation alongside |
| Whether something is true or just an idea | No way to tell from the answer | Documents can be marked as proposed or approved |
| If the answer changes tomorrow | Nobody notices | It's measured: the same question is re-evaluated against the current corpus |
| How it's checked | You have to trust the model | You open the file it cites and read it |
| If the model gets it wrong | It stays in the text, with no trace | The finding gets logged and stays searchable for the next query |
source for everyone
For the whole company
Each area asks about its own work, with a citation for every answer.
Leadership
Why did we decide to switch providers last year?
Operations
What's the actual procedure for onboarding a customer?
Finance
What services are we still paying for that nobody uses?
Sales
At what step in the process are we losing customers?
Technology
What breaks if we touch this service?
Legal
What terms did we sign with this provider?
Beyond code
It was born for technical teams. It delivers the most in product and technology, where all its tools work. The same cross-referencing of sources already answers business questions with data from every commercial platform.
Business questions, in the case
- Auditing cloud spend resource by resource
- Finding licenses and services still being paid for without being used
- Cross-referencing contracts, signatures and payments to see where a workflow gets stuck
- Reading the sales funnel with data the company already had and nobody had connected
- Every business service connected: CRM, e-signature, advertising, analytics, email, notifications, cloud and more
Product and technology
The team's code, tickets, PRs, chat and meetings. This is where its tools add value and where the case results were measured.
By the key your company already uses
Brain links sources by the identifier that shows up in all of them: a ticket, a case file, a client, a contract. An initial interview shows what information exists and how it connects.
Start small
A single area or a single source is enough to try it; the rest gets added later.
a person approves
In development
Where it shows the most: your tech team.
Understanding legacy systems
What a vendor who's no longer around built, explained through its tickets and changes.
Estimating with your own data
How long something similar took, how many systems it touched, how many people it needed.
Ship more, with control
The AI prepares the change. A person reviews and approves it: there's always a human in the loop, bringing their experience and knowledge to validate it.
Diagnosing incidents
When something fails, it cross-references alarms, deployments, tickets and changes to get to the cause.
Where it pays off
Situations almost any company will recognize.
You inherited someone else's system
What was handed over, finally readable.
Many roles, none sees everything
You ask once and it cross-references the sources.
You're short on specialists
An agent that knows your real system.
Auditing what you have running
And knowing whether something was left there on purpose.
How it works
Five pieces in a row and a ring of controls around the agent. Connectors that start by reading your systems, a knowledge base on your own infrastructure that rebuilds from the sources, and scoped tools that cite every fragment.
What your company already has
Connectors to your sources
They start by reading. They normalize, transcribe audio on your own machine, and read images with OCR.
SQLite + full-text search
A single file. Everything links up by the ticket key.
MCP server
Scoped tools: they start read-only and, when appropriate, get write access so the AI can start managing, with a person approving.
Your AI agent
Ask in natural language; it answers with citations.
the agent
How it's built
Brain isn't a separate application: it's a harness that installs into your AI.
A harness is everything that surrounds the model and governs how it works: the rules it reads at startup, the tools it can use, the procedures it follows and the controls that stop it. Brain installs inside your AI agent and adds four things: an MCP server with the query tools, skills with the procedures, rules your agent reads at startup, and hooks that block risky actions before they happen. That last part is the hardening: toughening the agent so that what must not happen doesn't depend on someone remembering. Your agent is still the same one; with Brain it works with your company's memory and within those limits.
What surrounds the model: rules, tools, procedures and controls.
Controls that block the risky action before it happens. They don't warn: they stop.
A harness with memory for AI agents. It installs into your AI; it isn't run separately.
It doesn't launch from an IDE: you open your agent in yourproject/ and it's already running.
picks which one
The tools
Each one for a type of question. The real ones, with a sample question.
Code history
5explain_changeWhy was the change in this ticket made?find_featureWhere is user sign-up implemented?find_similar_changesHave we built anything like this integration before?find_code_areaWhich part of the code touches payments?find_symbolWhere is this function defined?Documents and files
4search_docsWhat does the architecture documentation say about queues?search_confluenceWhat was left in the original project's wiki?search_jira_legacyWhat was requested about signatures in the first phase?search_mediaWhat was shown in the testing recordings?Conversations and decisions
2search_slackWhat was discussed about the provider's timeout?search_minutasWhat was decided in the meetings about the migration?Operations
3fallos_pipelineWhy did the deploy pipeline fail this week?historial_alarmaHow many times did this alarm fire, and how long did it take to recover?alarmas_sin_recuperarWhich alarms still haven't recovered?From zero to querying
Installs in minutes. The corpus will take as long as it needs, depending on the volume of information.
Install
~10 minutesOne command sets up the empty database, connects the server to your agent and activates the rules.
Load the corpus
depending on volumeA guided procedure loads the documents on its own and asks you for access to tickets, repositories and chat, one source at a time. What takes longest is the network.
Ask
from day oneFrom your AI agent, in natural language. When something isn't there, it says so instead of filling in the blanks.
Two ways to use it
You start on one machine. When the team needs the same memory, the corpus moves to a company server.
On your machine
One install per person, with their own corpus. It's how you get started and try it out.
- Installing takes minutes: one command. Loading and linking the data takes longer, depending on the volume, how well documented it is and how it's related
- Your memory on your disk: a local file, no server or cloud
- Your own data, cross-referenced: documents, repos, tickets, chats and meetings
- No extra keys: it uses the AI you already have
- Connects to the company server later, and your own memory stays alongside
- Core tested on Windows and macOS; Linux in continuous integration
One server for the whole team
A single database, loaded once, queried by the whole team. No duplicated data: everyone shares the same centralized information.
- Information loaded once for everyone
- Rules: what's approved kept apart from what's proposed
Tickets, code, chat, meetings, documents
The same as in the local version.
Connectors, database and query server
Searches and returns fragments with their source. It doesn't call any model.
no AI · no tokensHow it's fed
The same connectors and the same procedure as the local version. To add what's new, you run them again.
How each person connects
install.py --conectar adds the server to their agent; their local brain stays alongside.
Where the AI gets spent
In each person's agent, on their account. The server only searches: no model, no key, no tokens.
What's validated
Documents marked as proposed or approved, and the citation shows it.
Who gets in
Behind a tunnel or the private network; without a token it won't start outside the local machine. No per-person permissions.
What doesn't change
The same tools and the same citations.
distrust itself
The controls
A knowledge system fails when it confidently answers something that's no longer true. The controls make that error visible.
Every fragment with its source
And “not there” is different from “couldn't check”: a query that failed isn't an absence.
Rules that block
Deleting data, mass sending or writing without a log gets stopped before it happens.
Controls tested by breaking them
Every check was accepted only after watching it turn red with data broken on purpose.
Verification in a separate context
Subagents that re-measure the figures from the source, without seeing the reasoning.
What it reads is data
Whatever comes out of the corpus comes in marked as evidence. It mitigates prompt injection; it doesn't eliminate it.
It doesn't merge or deploy
The agent opens the change and a person approves it.
Controls are activated on each machine
The file that hooks them into the agent belongs to each installation and isn't versioned. The installer hooks them up in one step and, if even one points to a file that doesn't exist, it hooks up none: half a protection is worse than none.
The principles
Rules written in code, not in a manual.
A check that can never fail isn't a check
Every control is tested by breaking the data on purpose and watching it turn red.
Nothing gets deleted: it gets renamed
Not a table, not a file. Rolling back is a rename, not a restore.
Every claim carries how it was measured
And if the source changes on its own, when as well.
The structure is the backup policy
The product and your data don't share a folder
| Folder | Backed up? |
|---|---|
data/ what comes in | yes, the only thing that can't be recovered |
config/ your vocabulary | yes |
corpus/ the database | no, it gets rebuilt |
brain/ the product | no, it gets cloned again |
logs/ | no |
Upgrading is one command that tells you what the jump costs before it starts, and it can roll back to the previous version. Rolling back doesn't touch the corpus, and it tells you so.
at first glance
What's inside
What you see are the tools; behind them are months of work. All measured on the product repository, version 0.4.14, on October 11, 2026.
The 132 CLI tools, by family
Number of scripts in the CLI tools folder
Classified by each script's name prefix.
Where the 65,874 lines are
Lines of Python by product area
Plus 79 lines from a sample used in the security tests.
The 26 skills it ships with
Procedures the agent invokes on its own when the request calls for them, or that you call by name. All with the prefix placedium-brain-.
The catalog ones load on their own before claiming, measuring, deleting, touching infrastructure or trusting a permission, with the 166 cases that already happened.
Security
Audited against OWASP ASVS 5.0, with the “fail” results in plain sight. OWASP ASVS is the OWASP Foundation's open Application Security Verification Standard (owasp.org). It's a level 1 self-assessment, not a certification: it states what passes, what doesn't and what couldn't be checked.
55 requirements evaluated, across 12 chapters
Latest audit, version 0.4.14 · server and ingestion
Three audits, one per version since 0.4.12. The threat model declares 7 accepted risks, each with its requirement.
What runs on every PR and every release
Security controls in continuous integration and in the product
- Secrets: an in-house auditor on every PR, and gitleaks over the full history on every release
- Dependencies: osv-scanner and a software bill of materials (SBOM) generated on every release
- Code: static analysis with bandit on every PR, with a baseline that can only shrink
- Release gate: nothing ships with an outdated ASVS audit or with an unaccepted “fail”
- Server: off the local machine it won't start without an access token, unless deliberately enabled
- Corpus: everything that comes out of the database comes in marked as evidence, never as an instruction
- Boundary: no customer data gets into the repository: the .gitignore blocks it by pattern, and there's a commit check that activates in every clone
under your roof
Your data
The corpus lives inside your perimeter.
The full corpus is never sent to the AI
The model receives the fragments the search found for that question, and the answer says which ones.
The database and the server run on your infrastructure
On your machine or in your cloud account, with your logs.
The AI is your account
Brain doesn't sit between you and the model: you use your own AI agent and your own data agreement.
You decide what goes in
Through the direct chat connection, only the channels the bot is in come in, with no direct messages; if you load a full export, a restricted profile keeps them out of the answers. And there's an audit that detects personal data after loading.
What sets it apart
Inside, there's a database with text search. That's the easy part. Indexing documents takes an afternoon; what takes months is everything around the database: linking sources, reading what isn't text, knowing when something isn't there, and stopping the agent.
| A database with search | Brain | |
|---|---|---|
| What it returns | The documents where the word appears | The full chain: request, discussion, code change and outcome, linked by the ticket |
| Who searches | A person who has to guess the keywords | An AI agent that picks among several tools and cross-references sources on its own |
| What it can read | Whatever is already text | Also transcribed meetings, images via OCR and old exports nobody could open |
| When it finds nothing | An empty list, which reads as “doesn't exist” | If a source didn't respond, it says “couldn't check” instead of “not there” |
| What to trust | Everything carries the same weight | Documents are marked as proposed or approved, and the citation shows it |
| What surrounds it | Nothing | Controls that block, subagents that verify, and evaluation cases |
| How it improves | It gets reindexed | Every correction stays in the corpus, and the ones that matter become evaluation cases |
The ones that index your Drive
Cloud products that connect Notion (notion.com), Google Drive (google.com/drive) or Confluence (atlassian.com). Brain lives on your infrastructure and understands code, tickets and PRs, not just documents.
The ones that search by similarity
A vector database returns what's similar. Engineering questions call for what's exact: a ticket, a file, a service.
The personal wiki
Great for one person. A company is missing curation, controls and team access, which is what Brain adds.
Generic AI
It knows how to code and write, but it doesn't know your company. Brain gives it the memory and context it needs to understand it, with a reference to every source.
Where it comes from
The idea of a second brain for AI isn't ours. Bringing it to a company, with controls, is.
The personal second brain, popularized by Tiago Forte, author of Building a Second Brain (source).
Anthropic publishes an open standard for connecting AI to your data, now adopted by ChatGPT, Gemini and Copilot (source).
Andrej Karpathy, OpenAI cofounder and now at Anthropic: what matters is giving the AI just the right information (source).
The “company brain” becomes a startup category.
Brain by Placedium, built while operating a real fintech, with controls so a company can trust what it answers.
A database compiled from raw sources is enough for one person. A company also needs to link code, tickets, chat and meetings, mark what's validated, serve the whole team and stop the agent before it does damage.
The case · a 4-year-old fintech
The longest-tenured person on the project was leaving, the one who had been there since it was built. Operations carried on. The corpus was built in about six weeks, with no cloud infrastructure. Then the same system went from answering questions to sustaining delivery.
Product releases that reached production, by month
2026 · January through October 11 (* October, partial month)
Note: January–June average: 25 releases in 6 months, 4.2 per month; September (33) is 7.9 times that average. Through July, releases announced in the production channel; from August on, deployment batches closed in production according to their own log. Includes partial ones. Batches are smaller now: this counts deliveries, not volume; the comparable volume is PRs per business day.
What kind of system it handled
6.8 vs 4.5 compares the capacity of two ways of working, not individual performance: an agent with a single person reviewing, versus a team of 2 developers in Jan–Jun. Steady-state pace over a 3-week series, cut off on Aug 21, 2026; recalculated with each closed month.
release a day
The pace
Thirteen weeks in a row, and almost one new release a day.
Hours of work per week
Wall-clock time with activity in the sessions, case and product
Week of Jul 13 is incomplete: that's where the logs start.
Published product releases
Cumulative since the first one, on Aug 28, 2026
42 releases in 44 days. Each one says what it brings and why.
Built while operating
It's not hype. It's what was learned after getting it wrong many times.
The first real test case was running a fintech's platform with AI, every day since July 2026, turning every mistake into a rule.
At first, almost no error failed visibly. They came out green: a control that measured the process and not the data, a backup that went from 31 GB to 404 MB and said OK, a platform that returned zero when it didn't mean zero. Those cases are now a failure-mode catalog, in 10 pieces, that ships with the product.
Hours: wall-clock time with activity in the session logs, without double-counting parallel sessions; the available logs start on Jul 13, so it's a floor. Product measured on Oct 11, 2026 against version 0.4.14.
you find out
The limits
What Brain doesn't do.
It needs an AI agent
Brain is the memory and the controls. Without an agent it can be used from the command line, but it loses its point.
The corpus depends on you
Ten minutes for a local install on one machine. Loading the sources and linking them takes longer: it depends on the volume of information, how well documented it is and how it relates across systems.
A single source gives you a very good search engine, not a brain
The value shows up when you cross tickets, code, chat and meetings.
It doesn't know what isn't in the corpus
A stale corpus answers confidently about something that changed. It doesn't update itself: you have to reload it when the sources change.
Controls are hooked up per machine
The installer does it, but in an uninstalled copy of the repository none of them run.
Controls don't guarantee zero errors
They make one class of error visible. If the question is badly framed, everything can still come out green.
Using Brain takes a change in mindset
You can own a Ferrari and not have a trained driver: the best tool in the wrong hands gives bad results. You have to change the way you think and organize, and use Brain as part of the company's process: update, correct, train.
It depends on the quality of your sources
If you don't have documentation, if it's poor, or if the knowledge to link the sources is missing, the result will be poor. But all is not lost: when it doesn't exist, that's the moment to reorganize and start building it.
Frequently asked questions
What people ask before installing it.
Do I need an API key?
Not to install, load or query: Brain uses your AI agent. You only need one if you want to run the evaluation layer that uses a model as a judge, which sends those cases to the model with your key.
Why doesn't it use a vector database?
By design. Real questions ask for exact identifiers: a ticket, a file, a service. That's where full-text search works better.
Does it tell me whether a figure is correct?
No. It tells you whether it can be audited: where it comes from and whether it stopped matching its source.
Does it decide or merge for me?
No. There's no automatic merge and no writing without your approval.
Can I load just the documents to try it out?
Yes, and it's the fastest way to see it working. With a single source you'll get a very good search engine; the rest shows up when you add tickets, code and chat.
And it keeps learning
The initial load is just the beginning. Your Brain grows with use: every new piece of data, every correction and every rule make it more useful for your company.
Living information
You keep loading what's new: tickets, chats, meetings and changes go into the memory and become available to everyone.
Cross-references that used to be unthinkable
Knowledge and systems in the same query: analyses, statistics, projections, audits and sales opportunities that used to require pulling data together by hand.
Rules that add up
Every rule and every correction you add, your Brain takes in and applies from then on.
The repetitive gets automated
Whatever gets done twice becomes a procedure the AI runs and a person approves.
Load.
Ask.
“AI won't replace you. Someone who uses AI better than you will.”
Try it on your own history before they leave.
The repository is still private. Early access includes hands-on support with the first load.