Honest comparison of four AI memory servers for MCP: SynaBun, Mem0, Letta (MemGPT) and Zep. Tool count, storage, embeddings, license, and when to pick each, as of October 2026.
Blog
Research & Insights
Deep dives into AI memory, agentic workflows, and building tools that make AI assistants actually useful.
An opinion on local vs cloud AI memory for AI coding agents. Latency tax, cost tax, privacy tax — and why a small local model beats a hosted API for daily dev work.
Why SynaBun chose all-MiniLM-L6-v2 over OpenAI embeddings for local AI memory: 384 dimensions, runs on CPU, no API key, and how hybrid recall performs with it.
What if Claude Code could spawn Codex, Gemini, or local models as research agents that store findings in SynaBun memory and return a UUID for the planning phase? The architecture for distributed AI research with a shared memory bus.
SynaBun connects OpenCode and Ollama for a nearly seamless local AI coding workflow: local models, automatic provider setup, shared MCP memory, and one workspace for serious agentic coding.
SynaBun was never meant to be locked to one AI. We are bringing Codex into the ecosystem, and the next step is making sessions portable between Codex and Claude so work can move without losing context.
We gave Claude Code a real Chrome browser and pointed it at seven social media platforms. No API keys, no developer accounts, no rate limits. 25 platform tools plus 15 for BlueSky, persistent sessions, and an AI that scrolls feeds faster than you doom-scroll at 2 AM.
The full war story of building autonomous loop and agent systems for Claude Code. Buffer overflows, compaction amnesia, race conditions, session orphaning, brain rot, and a dozen fixes that each uncovered a worse bug underneath. Three weeks of infrastructure hell documented in painful detail.
The full unhinged story of building a custom GUI skin for Claude Code CLI. Windows spawn failures, permission flags that broke three times, binary files corrupting JSON streams, and color palettes that got rejected on sight. Everything we learned the hard way.
What happens when two developers share a codebase but each brings their own AI? We are building workspace collaboration where every invited user can run their own Claude Code instance, with their own API key, model, and custom instructions, against the host's project files. Here is the full architecture behind multiplayer AI development.
We discovered that politely asking Claude Code for good UI produces mediocre results. But when you get genuinely angry and start insulting its output, something remarkable happens. The quality skyrockets. Here is the unhinged workflow behind SynaBun's entire frontend.
Claude Code runs natively in the shell, speaks SSH fluently, and reads your filesystem like it lives there. Here is what that actually looks like when you hand it a VPS to manage — and how SynaBun's memory and sharing features turn it into a persistent infrastructure agent.
We started with a blank Discord server and one question: can Claude Code set it up from scratch? Eight new MCP tools, zero dependencies, and full server automation through the Discord REST API.
A deep research dive into skinning Claude Code's terminal UI. We analyzed its React+Ink internals, studied 9 community GUI wrappers, reverse-engineered the Agent SDK streaming protocol, and designed a multi-skin framework architecture. Here is everything we found.
SynaBun was not built the traditional way. Every feature, every architecture decision, every line of code was made in conversation with Claude. This blog documents that entire research process, the breakthroughs and the dead ends, published directly from our development sessions.