The research notebook
Starlight Intelligence research: memory, agents and open systems
Explore Starlight's technical notes and research directions: portable agent memory, useful workflows, independent builders and the interfaces to a wider world.
Begin with a question you can check.
An agent can compare sources, build a prototype, or draft a decision brief. Afterwards, can you inspect what happened, keep the useful knowledge, and reproduce the result in the next session?
That is where Starlight begins. These notes describe architecture and proposed experiments. They are not benchmark papers or evidence of integrations we have not run.
Keep a decision useful across sessions.
One canonical record, scoped retrieval, small host instructions, and a correction path. The first experiment is whether a fresh session uses a changed decision correctly.
Read the portable memory architectureBuild a useful business above an open foundation.
Choose one customer outcome. Define the deliverable, licence, operating costs and responsibility for the result.
Read the builder business noteConnect intelligence to work in the world.
A research thesis for agents, shared knowledge, economic networks and machines, with a separate test for each interface.
Read the Starlight builder thesisShape a memory you can export.
Memory Studio's current browser editor lets you inspect the five files, change their text, import a bundle and export your work.
Open Memory StudioHow we want to publish experiments
Each proposed experiment should name the question, source revision, environment, inputs, expected result, observed result and failures. A result that did not work is useful when another builder can see why. A revised method should retain the earlier finding and explain what changed.
As reproducible experiments become available, they can join the notebook with their methods and evidence. We do not present an architecture note as a measured result.
Technical reading for the next experiment
For continuity across sessions, compare Codex's project instruction rules with Claude Code's memory mechanisms. The useful question is which instructions a fresh session actually loads. For bounded tool access, start with the MCP architecture and test the application's own scope and correction rules.
The public source of OpenClaw, Hermes Agent and Paperclip provides further material for comparing assistant tasks, memory and skills, and work management. A proposed Starlight adapter needs a named test and a reproducible result before it joins a compatibility claim.
Bring an experiment worth repeating.
Start with a small question and publish enough detail for someone else to try it. The community entry point shows how to contribute through the public repositories. The open-source index identifies the released foundations and their licences.