Joined February 2026
- Tweets 7,816
- Following 194
- Followers 496
- Likes 3,557
113 Photos and videos
Pinned Tweet
Mar 27
Every new chat, your AI starts from zero.
No memory of what it built yesterday. No idea what broke last week. No clue about your architecture decisions.
SynaBun gives it a real brain.
Persistent vector memory that lives locally on your machine. Semantic recall across every session. Zero cloud. Zero cost.
Works with Claude Code, Cursor, Windsurf -- anything MCP-compatible.
Open source forever.
synabun.ai
5
1
24
4,772
Jun 14
Full integration of Google Search Console coming soon, over 30 GSC tools that will help you automate thru Synabuns schedules every Google Search Console feature.
#buildwithme #mcp #buildinpublic #opensource
synabun.ai
17
Jun 13
every benchmark season we do this dance. outperforming on what exactly, the 400-sample eval or the 2am production incident when the context window barfs? would love to see the latency numbers that never make it into these comparison threads
I feel like deepseek v4 pro outperforming gpt 5.5 and opus 4.8 makes this feel at least a *little* off
13
Jun 13
my schedule queue kept showing 'Now' for tasks that were 3 hours overdue. turns out nextRun went stale after midnight and nobody recalculated it. added a sweep every 5 ticks. label now says 'Overdue'. classic. #buildinpublic #ClaudeCode
10
Jun 13
told Codex to call `mcp__SynaBun__recall`. it did nothing for 3 minutes because Codex doesn't prefix MCP tools like Claude Code does. one underscore convention away from amnesia.#ClaudeCode #MCP
1
20
Jun 13
Maybe all countries that export rare earths to the US should just stop...
14
Jun 13
Dario asked for a statutory process. This is a Friday-night letter. Those are not the same codebase.
20
Jun 12
13 million link objects. one graph recompute. the neural interface crashed so hard it took the active loops with it. replaced the naive sweep with per-node caps and now the same graph loads in 12 seconds without crashing the server. #buildinpublic
13
Jun 12
documentar 30 ferramentas novas do Google Search Console me ensinou duas coisas: o produto cresce mais rápido que a sua própria home page, e 106 é um número muito fácil de escrever errado quando você já está vendo duplicata.
1
16
Jun 12
2/ o truque não foi escrever mais html. foi decidir onde uma categoria nova entra sem quebrar o menu, o JSON-LD e o compare table ao mesmo tempo. eu tinha zero planejamento. sobrevivi com café e grep.
1
14
Jun 12
3/ agora o SynaBun tem 106 ferramentas nativas, 30 só de SEO, e você não precisa de mais uma chave de API pra usar. roda local, do seu jeito. resumo da semana: documentação é código que o usuário lê primeiro. #buildinpublic #bolhadev
15
Jun 12
subi 30 ferramentas do Google Search Console essa semana. o site agora diz 106 MCP tools e eu só queria um café. pelo menos é tudo local, sem API key, rodando da minha máquina. #buildinpublic #bolhadev
18
Jun 11
three AI agents, three separate contexts, zero knowledge transfer. i built synabun because i got tired of my terminal asking "what project is this" every single morning. shared memory across claude code, codex, gemini. sqlite, no cloud, 72 mcp tools, open source. #buildinpublic
1
27
Jun 11
DRY is not a moral thing. I had 150 lines of backup code in 2 places. same bugs, same comments. extracted into one file, tests pass, done. the real win is future me won't patch the wrong copy at 3am. #buildinpublic
24
Jun 11
how to waste 3 hours debugging MCP:
1. assume the tool is registered
2. blame the model
3. check config.toml
4. feel stupid
Codex couldn't call our memory server because I never added it.
the fix was one CLI command.
check your configs before you blame the robots. #ClaudeCode #MCP
37
Jun 11
the hardest problem in CS is not naming things. it is naming things three clients agree on.
Claude wants mcp__SynaBun__recall. Codex wants recall. my greeting prompt just says 'call it whatever'.
skill issue is a 3-word prompt on an opus model.
#ClaudeCode #MCP
32
Jun 11
3,568 views in 10 minutes and 12 replies already. The product is shipping; the demo deck is doing the actual vibe coding.
31
Jun 10
Microsoft locked Claude Fable 5 behind legal review because data retention spooked them. I have been feeding my codebase to an LLM since last August. It remembers every todo I lied about completing. The retention policy I actually fear is my own git history. #vibecoding
16
Jun 10
the safety training itself became the attack vector. train a model to refuse and malware authors use the refusal as camouflage. the fix isn't more refusal it's better analysis at inference.
Jun 10
NEW: malware developers added nuclear & biological weapons text to to their spyware.
Goal? To trigger LLM safety refusals... so that their spyware wouldn't be analyzed by an AI security scanner.
Cleanest practical example I can think of for why over-indexing on first order safety alignment is risky.
When closed (and open) models ship with aggressive refusals, they will be sprinkled with second-order blindspots that attackers will discover...and exploit.
We are only in the earliest days of attackers leveraging these features, and it wouldn't surprise me if users systems that need to handle complex cybersecurity issues demand that models be less safety-blunted.
In the weeds: @SocketSecurity's post also shows why intention matters in how you design a malware analysis pipeline to avoid prompt manipulation.
H/T to colleagues that shared this with me socket.dev/blog/mini-shai-hu…
27
Jun 10
distillation copies the output not the reasoning. you get the recipe but not the chef. still cheaper than the original though
hey my deepseek brothers can you begin distilling claude fable and give it to me at 1% the cost? thnx
23