Project memory that survives between sessions.
Every coding agent session starts cold. Last Tuesday’s naming rule and bug walk-through were in the chat, and the chat is gone. The next session starts blank, and you explain them again.
The model proposes; the system decides.
Ask practitioners what breaks in agent systems and the answer is rarely capability. The failure is quieter: the model decides, and nobody can tell sound from fluent.
The model itself, analyzed: families, architectures, tradeoffs.
Qwen 4: confirmed facts, the 27B ladder, local fit.
Alibaba confirmed at Apsara that Qwen 4 is in training, no date attached, while community timing signals point at a launch in weeks rather than months.
Five-layer knowledge hierarchy for agent memory.
An agent with one flat memory file does not forget things; it cannot stop remembering things it was not asked about. The fix is a tiered context where each knowledge layer enters through a different trigger.
Open source AI, grouped by the job
GitHub’s monthly trending page works for a language, because a language filter already defines the set. Open source AI asks a different question: the work spans five languages, and the useful cut is the job a repository does.
Persistent Hermes memory on your own hardware.
A self-hosted AI agent needs more than a long context window. It needs a way to retain useful facts, retrieve earlier decisions, and revise what it thinks it knows as projects evolve.
A software factory built from agent skills
AI coding agents can already write functions, modify repositories, run tests, and open pull requests. The harder problem is getting an agent to follow a repeatable engineering process before, during, and after the code is written.
Self-hosted agents that research beyond search
Deep Research has become its own category of software, not just a model pointed at a search box. This article compares twelve self-hosted systems and the research architectures behind them.
The 2026 sweet spot for open LLMs: around 30B.
In September 2026 the open-model efficient frontier sits between 25B and 34B parameters: near-frontier agentic work on one 24 GB GPU, at a fraction of 70B-class hardware.
Local Hermes memory with controlled writes.
Mnemosyne is a local-first memory provider for Hermes Agent, storing working memory, structured facts, temporal data, and episodic history in local SQLite — no hosted service, no mandatory network calls, and unusually granular write control.
When remembered conclusions become new evidence.
Persistent memory turns an agent from a re-explained tool into one that carries context forward — but it opens a failure mode stateless chat avoids: an interpretation can become memory, retrieved as fact, and justify a stronger version of itself.
No reject state. Here's the workaround.
An agent that proposed “move persistence into a shared library” six months ago will happily propose it again unless something durable says it was already rejected – and OpenSpec has no built-in state for that today.
Specs as deltas, not a 40-page PRD.
OpenSpec is a free, open-source CLI from Fission AI that gets you and your coding agent to agree on a change in plain Markdown before any code gets written, without the phase-gated ceremony of heavier spec-driven frameworks.
A safe cutover beyond the one-line import
Migrating an AI assistant is not the same as copying an application config. The hard part is preserving identity, memory, tool behavior, scheduled work, and messaging access without two gateways acting as the same bot.
When llama-server beats Ollama
Ollama and llama.cpp are often compared as if they were rival inference engines. The real choice is between a managed model service and a toolkit you operate directly.
