Conversations & corrections
Start with your recent Codex conversations.
From experience to better agents
Turn past conversations into lessons your agent can use. Review proposed improvements, reuse what works, and keep your goals, results, and feedback connected.
Send one prompt to your agent. Your own Codex. A clearer direction.
A better playbook, one lesson at a time.
Start with your recent Codex conversations.
Keep the source behind every suggestion.
Review first. Save the changes you trust.
Reusable guidance. A clearer direction.
Research reference · ICLR 2026


LuLu helps your agent extract lessons from past work, keep the evidence, and carry approved guidance into future tasks.
Published reference benchmarks · Gemini 2.5 Flash · compared with No Memory
WebArena success rate
+8.3 percentage points40.5% → 48.8%
684 tasks · Table 1
Fewer interaction steps
14.4%9.7 → 8.3
WebArena average · calculated from Table 1
SWE-Bench Verified resolve rate
+4.6 percentage points34.2% → 38.8%
Software engineering · Table 2
Source: Ouyang et al., ReasoningBank, ICLR 2026. Read the paper ↗ · Research repository ↗ · Apache 2.0
From conversations to outcomes
A goal is more than a task. Keep the purpose, the work, and your feedback connected as your understanding evolves.
Use recent Codex conversations to uncover goals, retries, and corrections. Review improvements with evidence from the work that produced them.
Buildings are goals. Roads connect the work. Explore your city, open a goal, and see exactly where things stand.
Read your agent's results, review proposed guidance, and approve the changes you want to keep. Give the next round a clearer direction.
01 / From real work
Choose recent Codex conversations. LuLu helps uncover retries, corrections, and useful patterns, with short evidence excerpts you can review.
“Run the required checks before releasing the change.”
State what passed, what was skipped, and what still needs attention.
Evidence attached · Awaiting review02 / Work you repeat
From a coding session to a research task, keep useful corrections within reach. These examples show the kinds of guidance you can review and reuse.
Connect your own Codex ↗03 / You stay in control
Read the evidence, edit the proposal, and approve what deserves to become guidance. Keep useful changes as a Prompt, then confirm any update to your project instructions.
Use the source evidence to decide which checks belong in your project guidance.
Let your agent get you started
Install LuLu, connect your account, and run your first analysis. Your goals become a city you can explore.
macOS / Linux · Codex installed and signed in
Send this prompt to your agent
Please install and connect Lulu on this computer: https://lulucli.com/proactive. If it is already installed, check and restore the local connection. Otherwise install using https://lulucli.com/proactive/install.sh and connect to https://lulucli.com/proactive. Open the sign-in page if needed, then link this computer automatically. Finally verify that this computer is connected to the current account and website.
curl -fsSL https://lulucli.com/proactive/install.sh | shRaw conversation history stays on your computer. Selected excerpts are analyzed by your configured Codex model service. LuLu saves goals, results, and the evidence needed to keep them grounded.