Vitalii Cherepanov

Founder of BrainCore, a brain on a spiking neural network (SNN) with no LLM inside. Author of total-agent-memory, memory for AI coding agents. Go · AI

[Vitalii Cherepanov @vbcherepanov]

Everyone is shipping "memory for AI agents" right now. Almost nobody publishes numbers.

Here are mine, with the harness so you can rerun them:

LongMemEval R@5 — 95.1%
LoCoMo — 0.607
BEAM-1M — 0.448

100% local, no API key. 🧵

[Vitalii Cherepanov @vbcherepanov]

Act meets draft? You get draft. A connector set to draft can never act, whatever the agent asks for. #aiagents #llm

[Vitalii Cherepanov @vbcherepanov]

Embeddings place text by meaning. Your error code has almost none, so a lookalike code wins. Hybrid search fixes it. #vectorsearch #rag

[Vitalii Cherepanov @vbcherepanov]

Approve or deny is not the same as editing a tool call's arguments before it runs. I built LOOPmKit, so this is not neutral: a sourced comparison with n8n, CrewAI, Dify and LangGraph. https://t.co/kv8oGwgp1m #AIAgents

[Vitalii Cherepanov @vbcherepanov]

Agents that call each other are a tangle. Agents that wake on records are a system. One job each. #aiagents #llm

[Vitalii Cherepanov @vbcherepanov]

GPT-5, o3 and four siblings lose API access Dec 11. If changing your agent's model takes more than one env variable, start now. #aiagents #llm

[Vitalii Cherepanov @vbcherepanov]

An instruction given once at session start is not enough. ~1 memory save in 30 sessions proved it. Remind the agent at the moment of work. #aiagents #llm

[Vitalii Cherepanov @vbcherepanov]

LOOPmKit: full product demo of a self-hosted AI agent builder https://t.co/asDIgLxLgL via @YouTube

[Vitalii Cherepanov @vbcherepanov]

Rewrote BrainCore's FFT with only the Python stdlib. 22x slower. Floats off by up to 1.56e-8. Neural input: unchanged. #fft #python

[Vitalii Cherepanov @vbcherepanov]

Every run green. Dashboard healthy. Data two days behind. Success per run says nothing about whether your pipeline keeps up. #aiagents #observability

[Vitalii Cherepanov @vbcherepanov]

SIGSTOP freezes a process. It saves nothing. We learned that after ~6 hours of brain simulation vanished. #checkpointing #neuroscience

[Vitalii Cherepanov @vbcherepanov]

One subagent built a whole site with 260,269 tokens that never touched my main context. For a one or two file change, I skip subagents entirely. Here is where I draw the line: https://t.co/DNXa6fAQvE #ClaudeCode #AIAgents

[Vitalii Cherepanov @vbcherepanov]

Auto-recall tuned to miss nothing will happily inject a 2003 pager website into your screenshot question. Every stray record costs tokens on every prompt. #aiagents #llm

[Vitalii Cherepanov @vbcherepanov]

A green test can mean the output bent to fit it. My agent turned Gmail into Google Mail to pass an overly broad rule. #aiagents #llm

[Vitalii Cherepanov @vbcherepanov]

No complete equations for mGluR6 → Go → TRPM1 turned up. So our digital brain didn't guess. A blank beats a made-up number. #braincore #neuroscience

[Vitalii Cherepanov @vbcherepanov]

HTTP 200 means the server answered. Not that you got data. My script saved 4 firewall block pages and called it a perfect run. #webscraping #dataquality

[Vitalii Cherepanov @vbcherepanov]

Your coding agents can't talk to each other, so you copy-paste between them. I fixed it with one open protocol (A2A 1.0) and five rules for when one agent may interrupt another. https://t.co/Q6mA8R7ch6 #AIAgents #A2A

[Vitalii Cherepanov @vbcherepanov]

A test that compares with a dependency is testing the version of that dependency. SQLite 3.45 rounds 23:59:59.9996 into next year; 3.50 truncates. #sqlite #testing

[Vitalii Cherepanov @vbcherepanov]

Replacing a dependency doesn't count until it reproduces the old answer bit-for-bit and the tests stay green. 18 of 54 executables to go through. #golang #testing

[Vitalii Cherepanov @vbcherepanov]

launchctl bootout stops a job until the next login. My agent learned that when a reboot revived a service its memory still called "stopped". #aiagents #macos

[Vitalii Cherepanov @vbcherepanov]

Deleting a wrong fact from your agent's memory doesn't stop it from saving it again. Proposed fix: tell the agent 'revived: true'. #aiagents #llm