Desk with a laptop, an open notebook and a pen

· Translated from the Spanish original

Artificial intelligence

A5 is a desktop assistant that remembers your context and proposes the next step

A team presented A5, an agent that keeps context across sessions, detects changes in your sources and suggests tasks. It uses your own model API.

What happened

  • On October 23, 2026, Honglin Guo and four other authors presented A5 in a Hugging Face community article. They define it as a proactive personal assistant that understands you better the more you use it.
  • Unlike agents organized by conversation, A5 keeps a long-term context about you and your projects and reuses it: it applies preferences from earlier trips, for example, to a new trip.
  • It detects changes in connected sources, adjusts ongoing work and proposes next steps through a component called Proposal Engine.
  • It installs as a desktop app. It needs a model API in the Anthropic Messages or OpenAI Chat Completions format, with streaming and tool calling.

Why it matters

  • Today the user often carries the work of moving information from one conversation to another. A5 proposes that the agent do it.
  • Proactivity has a cost: an assistant that acts without being asked needs permissions and sound judgment about when to step in.
  • Accepting third-party models by API lets you choose a provider, but quality will depend on the model connected.

The number

2 formats. API styles accepted: Anthropic Messages and OpenAI Chat Completions.

Context

The text describes an architecture with a Harness, the technical base for collaboration across sessions, context management and background execution. Context updates enter before the next call of the ReAct loop so the agent can revise its plan. Web search and page fetching use free services by default. The demo examples cover project management and trip planning. The authors cite earlier work on agent memory, including MemGPT and A-MEM.

What’s next

  • The authors plan to release A5 as open source on GitHub «soon», with no date.
  • They name continuous learning from the Harness as a next step.

Bottom line

Memory is the underlying problem: MemT-1.0 puts memory inside the model and promises 66 times less storage. Remembering also carries risks, as shown by an AI assistant recovering up to 54.6% of secrets after a change of topic. There are other personal agents in specific domains, such as the personal health agent Nature published, and for companies, MCP connects sales agents with AI assistants.

Sources


Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

Related notes

← All notes