Instant agent memory

The product pitch is simple: your agent should remember useful work immediately and retrieve it before the next answer.

Agent memory should feel like a reflex. Before an agent answers, it searches what the team already knows. After it makes a useful decision, it writes the durable part back. The next turn starts with that context available instead of buried in a transcript.

GPU Brain packages that loop as hosted memory. The agent gets tools for search, page reads, writes, tags, typed links, backlinks, and timelines. Humans still get markdown pages they can edit and audit.

The operating loop

• Search before answer: retrieve relevant project, person, company, and decision pages first.
• Read the source: open exact pages instead of relying on vague search snippets.
• Act with context: use REST, CLI, hosted MCP, or local MCP from the same account.
• Write after decisions: save what should survive into markdown, tags, links, and timeline entries.

What came from Gbrain

The larger Gbrain project proved the pattern: brain-first lookup, typed relationships, timelines, collections, and agent-operated workflows. GPU Brain brings the parts that matter for hosted instant memory into a small Go service with a web UI, CLI, and MCP server.