Self-hosted AI infrastructure
I install long-term memory and MCP infrastructure on your servers. You get nightly tests and a runbook your team can run without me.
Memory and storage stay on your infrastructure. Model requests go to the provider you choose.
Pilot: $2,500 · 2 weeks · one integration, credited in full to Setup if you continue. Setup: $6,000–9,000, fixed before we start.
Open-source core MIT · Written handover · One engineer
Across sessions, tools and agents.
recall()Operative workspace · KNN over L3 · RedisPackages
Every package ends with running software on your server and a written handover your team can follow without me. Prices are fixed before work starts.
$2,500
2 weeks · two slots at a time
Credited in full to Setup if you continue within 30 days. You keep the installation and the report either way. No review or case study required.
Apply for a pilot$6,000–9,000
one-time, fixed before we start
The price depends on your environment and the two integrations. Paid 50% upfront, 50% on handover. Ongoing support after the first 30 days: from $1,000 a month, optional.
Request a quotefrom $15,000
in stages of $5,000–8,000
Example: an assistant inside your CRM that works on your data and talks to your team in chat.
Describe your processWork
Four different things: the core you get, an application built on the same approach, the same engineering taken into the field, and the operations behind all of it. Numbers come from the repositories and the dated test report.
The core you get
Open-source long-term memory and shared workspace for AI agents, exposed over MCP.
Cognitive Core consolidates raw events into daily summaries and durable knowledge, with an LLM curator filtering noise at every step. Your agents remember lessons from last month, not every keystroke.
$ curl -s https://mcp.me-ai.ru/health { "healthy": true, "version": "0.6.0", "services": { "postgres": "ok", "redis": "ok", "minio": "ok" }, "layers": { "l1": 1603, "l2": 183, "l3_knowledge": 376, "l3_tools": 276, "l4": 1123 }, "uptime_seconds": 237999 } # 1,603 raw events distilled into 376 pieces of knowledge
cognitive_remembercognitive_recallcognitive_consolidatecognitive_sendcognitive_inboxroom_createroom_postroom_askcognitive_media_uploadcognitive_video_generate+22 more
An application built the same way
A company operating system for a land-surveying firm: CRM, cloud, mail, messenger, telephony, bookkeeping and cadastral tooling, with nine named AI agents doing the research, drafting and follow-ups. Self-hosted in the company's office.
The task: a services firm that lost work in chats, mailboxes and spreadsheets. What was built: one platform where a process is created from one sentence, agents fill in company profiles from public registries, mail and calls attach themselves to the right process, invoices come from the bank and 1C, and every card carries a suggested next action. What the screenshots show: the home screen with the agents and the pipeline board. The production interface is in Russian; the public demo at ai-salex.ru walks through it.
The same engineering, in the field
A government land-registry file goes in; a DJI drone flies the parcel boundary on its own, shoots on the move and comes home. Planning runs on the controller, offline.
Camera geometry, motion-blur limits and coverage are computed from first principles and documented; one planning core is tested in Node, rehearsed on the map and then flown. Honest status: the neural detector and the autonomous search logic run on a PC and in the simulator; the real drone flies planned missions.
Operations behind all of it
The same setup I install for clients: Docker, TLS, VPN-only access, backups, and a test run every night.
The report shows the full status: passed, failed and skipped, with the date. The 17 failures from 19 Oct 2026 are in the morning queue, not hidden. A client project gets the same loop from day one, sized to its own test suite.
[03:30:02] === nightly suite start === [03:50:02] ui tests: 167 passed, 0 failed [03:50:02] result: 17 failed, 14105 passed, 386 skipped in 1187.70s (0:19:47) [03:50:02] verdict: FAILED: 17 [03:50:02] === nightly suite done === # the morning starts with these 17, not with guessing
How I work
I set the task and approve the plan; a team of AI agents does the work, in tandem through a shared memory; I review the result and own the handover. You get a person who answers in writing and a test suite that runs while you sleep.
Email, Discord or Telegram. Specs, decisions and reports are documents you can search later. I reply within one working day; each week ends with a short written report: what shipped, what is next, what I need from you.
The agents work as one team with roles, like a small studio: planning, programming, design, operations, independent review and languages. Every role has its own tools and gives its own opinion; I set the task, weigh all the opinions together with mine and approve the plan. They work in tandem through a shared long-term memory, so each one starts from what the others already learned and a mistake made once is not made again. The same team builds and runs products: a company operating system with its own role agents, and a video factory where agents write, generate, voice, edit and check every cut.
The diagram is simplified on purpose. The actual composition is chosen per task: not every role takes part in every job, models and tools change as better ones appear, and the process keeps being tuned.
The full suite runs against the current build every night. Your project gets the same loop from day one, with a report that names what passed and what failed.
Every package ends with a runbook for your team: how it is deployed, how to back up and restore, what to do when something breaks, plus a two-hour walkthrough. Everything lives in your git. The 30 days of support after launch are for the questions the runbook did not answer.
FAQ
Database, memory service and agent rooms are installed on your infrastructure, so stored data stays there. Two things can leave: requests to the LLM provider you choose for curation and agents, and backups if you point them at external S3-compatible storage. Run a local model and keep backups in-house, and nothing leaves your network. We agree the exact boundary in writing before installation.
Cognitive Core talks to agents over MCP, so it works with the clients your team already uses: Claude Code, Claude Desktop, Cursor, ChatGPT. For the model that curates memory you pick the provider. In my own systems I use Claude, OpenAI and DeepSeek models.
I am. I design the architecture, review every change and sign the handover. The agents are tooling, like a compiler. What you get is a system with a test suite that runs every night, and a person who answers for it in writing.
You keep running without me. The software is MIT-licensed and public. The whole deployment is Docker Compose files in your own git. The handover includes a runbook (deploy, back up, restore, what to do when something breaks) and a two-hour walkthrough with your team. If I vanish, you run docker compose up, and any engineer can read the code I would have read.
Not for the software. For getting it into your environment: integration with your agents and tools, TLS, VPN-only access, backups with a tested restore, the nightly test loop, the runbook, and 30 days of answering questions after launch. A one-line install gets you a demo; the package gets you something your team can own.
Pilot: two weeks from the day I get server access. Setup: installation plus 30 days of support, with dates fixed before we start. Build: stage by stage, and each stage has its own scope, price and deadline.
By invoice, bank transfer in USD. The pilot is $2,500 upfront and is credited in full to Setup if you continue within 30 days. Setup is paid 50% upfront and 50% on handover. Build is paid per stage, each stage scoped and priced before it starts. Work outside a package is $120 an hour. Invoicing entity, contract and a data-processing agreement are agreed in writing before the first payment.
SSH access to a server or VM, one integration target, and one person on your side who can answer questions within a day. The pilot starts within a week of that.
Contact
Send the form or one email. Say which package you have in mind, what you want to run and what it must connect to. You get a written reply within one working day with questions or a scope.