Software developer · founder of BrainCore
Software developer since 2002. Founder of BrainCore. Author of total-agent-memory, memory for AI coding agents.
Vitalii Cherepanov
Software developer · Founder of BrainCore · Author of total-agent-memory
The story
I started programming in 1998, while still at school, and have worked as a software developer since 2002. The first years were automation jobs in Irkutsk: an engineer at Baikal-Global-Service and head of the automation sector at Sibinstrom. From December 2004 to June 2010 I served in the Russian Armed Forces (conscription, then contract) and kept programming in my spare time.
After the service I worked in Tikhoretsk as a software engineer and head of automation, and from 2012 as a sole proprietor: PHP, Symfony and 1C-Bitrix systems for e-commerce, payment integrations and CRM, with modules sold on the 1C-Bitrix Marketplace in 2015 to 2017. In 2020 to 2022 I was a lead technical programmer and then a software development engineer at DeMedia. Since 2022 I work under contracts for clients in Russia, the US, Australia and Serbia.
Since February 2026 I build open-source infrastructure for AI coding agents: total-agent-memory (a local memory server over MCP), a2abridge (an agent mesh on the A2A 1.0 protocol) and a2a-php-sdk. In 2026 I founded BrainCore, a human brain built in software as a biologically detailed spiking neural network.
Since August 2025 I live in Novi Sad, Serbia, and work as a registered entrepreneur.
Featured work
Open-source infrastructure for the AI coding agent era, plus production SaaS platforms delivered for international teams.
Distributed cognitive workforce runtime — the AISWARM + a2abridge + total-agent-memory triad.
Pure-Go orchestrator that runs Claude Code, Codex, DeepSeek and local Llama (LM Studio / llama.cpp) side-by-side on one codebase. Phase 6 closed: 33 internal packages, 8 binaries, distributed cognitive workforce runtime. KindPlanner expands tasks into 12 typed agent roles; auto_choose routes 30/30/20/20 across providers; a 3-tier verifier (command → regression-on-integration → LLM judge) gates merges; peer-helper consult (ASK_PEER) spawns cross-model helpers on demand; soft pause/resume protects dependency cascades; subtasks decomposed over MCP; tokens + cost tracked per provider. Workers run inside sandbox-exec / bwrap, each in its own tmux + git-worktree, with SQLite (modernc.org, no CGO) in ~/.aiswarm. Bubble Tea TUI and a Web UI with REST API show A2A and memory tabs live. Memory is a side-channel through MCP (total-agent-memory or any MCP memory server) — the runtime does not embed its own. Two open peers stand under it: a2abridge (A2A 1.0 mesh) and total-agent-memory (memory plane).
Persistent memory for AI coding agents — local or as a team server.
MCP memory server for Claude Code, Codex CLI, Cursor, Hermes Agent and other MCP clients. Version 14.9.0 (2026-10-13): hooks recall, check rules and save without a memory tool call; one warm daemon shared by every client (first recall 0.18 s instead of 1.51 s); LongMemEval-S held-out 88.0; 79 tools, CI on Ubuntu, macOS and Windows. 14.8.0 (2026-10-08): every recall tier measured, a graph tier that never fired and a representation tier that scored an arbitrary subset fixed, R@10 on the author's store 0.133 → 0.167 at a third of the latency. 14.7.0 (2026-09-30): whole records fill the context budget; AMA-Bench full test set 0.658 (self-reported, submitted to the leaderboard). 14.6.0 (2026-09-28): a company server with a dashboard, roles, onboarding and PostgreSQL; encrypted settings and privacy tools; a Hermes Agent plugin. 14.5.1 (2026-09-24): a record stays below its own later update. 14.5.0 (2026-09-23): no statistically significant difference from Mem0 Platform on LoCoMo and LongMemEval when both are graded on the same questions. 14.4.0: a save can retire the value it replaces (opt-in supersede), and writes at a million records no longer reload the vector cache. 14.3.1 stopped dropping fact updates as repeats and keeps project-scoped recall near 100 ms at a million records; 14.3.0 answers with the latest value of a fact that changed over time and can check contradictions with TypeSafe’s Jev. 14.0 added an authenticated server with personal, team and shared memory, attributed revision history and a team web interface. Hybrid retrieval on SQLite, temporal knowledge graph, Ollama / OpenAI-compatible / Anthropic providers. MIT, Python 3.11+.
A human brain in software, region by region — a biologically detailed spiking neural network (SNN). No LLM, no backprop.
Research project (founder: Vitalii Cherepanov) building a complete human brain in software, region by region, from published neuroscience. Brain simulation as a spiking neural network: leaky integrate-and-fire neurons with Rotter–Diesmann exact integration, Potjans–Diesmann cortical microcircuit columns. A model retina feeds area 17 (reads 4 letters at 94 %, preregistered rule passed on 20/20 seeds); a Meddis cochlea feeds area 41 (tonotopy on 20/20 seeds). Spike-timing-dependent plasticity (STDP), dopamine and sleep mechanisms are built; hippocampus in progress. No LLM, no pretrained weights, no backpropagation. Written in Go, bit-exact deterministic (SHA-256 over every spike), 89.7M synaptic events/s on a laptop CPU; the spike signal suits neuromorphic chips. Paper 1 frozen on 11 Sep 2026. Research stage, raising pre-seed.
Docker disk cleanup that knows which project each leftover came from.
docker system df says 84 GB are gone but not whether that volume can be deleted. Once a project's containers are removed, nothing records where the project lived on disk — and Docker cannot reconstruct that link afterwards. flotsam keeps its own SQLite index of project → directory, filled passively from compose labels on every run, so months later it can still say: this 8 GB volume belonged to legacy-crm, which lived in /home/v/work/legacy-crm, and that directory is gone. Read-only probes name what is inside a volume (postgres/mysql/mongo/redis), a layer graph gives an honest reclaim estimate, 18 ordered rules assign a risk level, and apply is a dry run by default. Single static binary, no CGO, CLI plus an embedded Vue web UI.
Rich two-line status bar for Claude Code CLI.
Real-time model, context usage, tokens, cost, duration, git branch, cache stats — without leaving the terminal.
Open A2A 1.0 mesh for Claude Code, Codex CLI, Cursor, Cline, Continue, Gemini CLI and Antigravity CLI.
Wraps each coding agent into a standard A2A 1.0 peer. Version 4.0 (2026-09-15) serves the JSON-RPC and HTTP+JSON bindings with SSE streaming and push notifications through the official a2a-go SDK v2.5.0. Local directory on 127.0.0.1:7777; single Go binary, MIT.
A2A 1.0.0 client and server SDK for PHP 8.4 — JSON-RPC, REST and gRPC.
Official A2A SDKs exist for Python, JavaScript, Java, Go, .NET and Rust, but not for PHP, and the PHP packages on Packagist target protocol 0.3.0. This SDK implements all 11 A2A 1.0.0 operations over JSON-RPC, HTTP+JSON/REST and gRPC, with SSE and gRPC server streaming, using message classes generated from the official a2a.proto. Implement one Executor interface to serve an agent, or call any A2A agent through a typed client. The official a2a-tck runs in CI on all three transports.
A2A 1.0.0 agents in Symfony: routes, security, Doctrine and Messenger.
Symfony integration for a2a-php-sdk. One route import exposes JSON-RPC, REST, SSE and Agent Card discovery; the agent is a Symfony service, and outbound agents are injected as named clients. Tasks are scoped by caller identity and tenant through Bearer tokens or Symfony access-token handlers, stored in files or Doctrine DBAL and scheduled through Messenger, with optional native gRPC on OpenSwoole, JSON logs and Prometheus metrics. Tested on Symfony 7.4 and 8.0 and against PostgreSQL 18.
Filesystem as a message broker — exchanges, streams, dead letters.
Persistent streams, consumer groups, priority queues, DLQs, retries and metrics. Everything you expect from RabbitMQ, Kafka and NATS — but it's a single PHP package and a directory tree.
Bitrix Docker tooling that finally feels modern.
Repeatable, multi-stage Bitrix dev environments — PHP 8.x, MariaDB, push, sphinx, scheduled agents — all behind a single Makefile.
Six llama.cpp-inspired optimisation patterns, measured in PHP 8.4.
A reproducible benchmark suite that translates llama.cpp ideas — mmap'd weights, flat dense buffers, value pools, table dispatch, generator streaming, columnar layout — into idiomatic PHP 8.4 with JIT, plus a naive-vs-optimised case study importing 100K CSV rows into Postgres. Docker + Make, PHPStan level 8, no cooking the numbers.
Multi-tenant loyalty & bonus platform with a PremiumBonus-compatible POS API.
A hexagonal Symfony 8 / PostgreSQL 18 loyalty backend: wallets with typed balances (gifted / promo / earned), a targeted-bonus allocator that honours per-item and per-cart caps with FIFO expiry, and a 4-step POS redemption flow (request → confirm → commit → cancel) that is idempotent by document number. Ships a Bitrix D7 module for e-commerce integration and full GitLab CI/CD to Docker Swarm.
AI-assisted English-learning app — a single Go binary, live at fluenta.club.
An English-learning app that ships as one Go binary with an embedded web UI over PostgreSQL 18 — the full A1→C1/C2 Roadmap (6 levels, 60 units). Speaking practice runs local Whisper for on-device transcription; Claude Haiku grades speaking and writing (score + breakdown + model answer). Works fully offline without any key — catalog, exercises, and SM-2 spaced-repetition flashcards are all built in; an optional ANTHROPIC_API_KEY unlocks AI grading and exercise generation.
IT interview-prep & e-learning platform with spaced repetition and an AI mock interviewer.
A Symfony 7.4 / Nuxt 3 learning platform covering Go, PHP, Symfony, algorithms, system design and cloud: quizzes, exams, an FSRS spaced-repetition engine, gamification, RAG-grounded answers and an isolated code-execution sandbox. The same markdown corpus (800+ topics) also compiles into printable interview handbooks.
Writing
Long-form pieces on memory, AI tooling, distributed systems and the boring backend work behind them.
BrainCore step 2: a human ear model with 1,000 cochlear sections and 32,000 fibres, a cochlear nucleus cell at 1.25 microseconds, letter-to-sound pairing at 9 of 9 on a fixed course. Unknown voices are still not rejected. Every number links to a verification file.
A reproducible round trip over A2A 1.0: a PHP client calls the official Python hello-world agent, and the official Python client calls a Symfony agent. Pinned versions, real output, a 401 for a wrong token, and the two install bugs that only a real composer require could find.
Six llama.cpp optimisations translated to PHP 8.4 with JIT, benchmarked from 1M to 1B records. Half the hypotheses were wrong: SplFixedArray loses on speed, mmap is 7× slower per call, match equals switch. The other half become survival tools at scale.
Analysis of Anthropic's experiment running 16 parallel Claude instances to build a C compiler in Rust — what infrastructure had to be invented around the agents because the primitives weren't there yet.
Production AI should optimise for zero confidently-wrong actions, not accuracy percentage. Abstain as a first-class outcome, self-tasking, and a cognitive runtime are the path to trustworthy autonomy.
Atomic units with lifecycle, strict mode with abstention, causal decision chains, AST-based code identity, internal git, scoring mechanics, and negative memory with rule engines — the recipe behind real cross-session memory.
Vector search is search, not memory. Three failure modes — chunks lose meaning, no structural distinction between facts, and no temporal validity — make RAG hallucinate confidently without knowing what it doesn't know.
Patterns that quietly accumulate debt: phantom abstractions, copy-paste-by-prompt, silent error degradation, context amnesia between requests, and decorative tests. Human oversight of architecture and immediate refactoring are the antidote.
Total Agent Memory — an open-source MCP server providing persistent knowledge across sessions. 32 tools spanning storage, self-improvement via error tracking, knowledge graphs, episodic recall, and skill assessment.
Seven architectural principles distilled from etcd's Go codebase, mapped onto everyday PHP/Symfony work: typed contracts, single-responsibility services, composable middleware, observability as architecture, facade APIs, traceable request paths, and explicit dependencies.
An architectural style for backend systems that combines Hexagonal, Outbox/Inbox, Observability, mandatory Feature Toggles, and flow visualisation.
Immersive VR tool for debugging and analysing complex microservice flows with Go, gRPC, and Unity VR — bottlenecks in seconds, request flows in real time.
Toolkit
Grouped by job, not by skill bar.
Now
BrainCore (getbraincore.com) — a human brain in software as a spiking neural network: model retina → area 17, cochlea → area 41, STDP. No LLM, no backprop. Paper 1 frozen on 11 Sep 2026.
total-agent-memory, claude-statusbar, a2abridge, phpfilebroker — six pinned OSS projects, all in active dogfooding.
Writing on Medium about AI tooling, memory architecture, and the infrastructure agents need to do real work.
Contracts for clients in Russia, the US, Australia and Serbia. Open to relocation and AI-tooling roles.
Connect
Looking for Backend and AI infrastructure work: async remote contracts or on-site staff roles. Open to relocation. Based in Novi Sad, Serbia, working with clients in Russia, the US, Australia and Serbia.