The three-minute, two-chat memory challenge

[Lians-ai]

Most AI tools feel smart until you open a new chat and have to explain the same thing again. This is a small, honest test of whether your agent can carry one useful fact across chats?and let you inspect, correct, and remove it afterward.

The three-minute challenge

First, add Lians using the free local MCP setup. It runs locally with SQLite and needs no Lians account or API key.

Use a harmless test fact, not personal or confidential information.

1. Remember in chat one

Remember that this test project uses Python 3.12 and pytest.

2. Recall in a new chat

Close that chat, open a new one, and ask:

What Python version and test runner does this test project use?

3. Correct it

Correct that memory: the test project now uses Python 3.13 and uv.

Then ask again in another chat. The current answer should not include the stale Python 3.12 and pytest combination.

4. Inspect and forget it

Show me the saved memory for this test project. Then prepare to forget it, but ask me for confirmation before deleting anything.

Confirm only after your agent shows what it intends to remove.

Your score

• 1/4: the fact survives a new chat;
• 2/4: recall returns the small relevant fact instead of a conversation dump;
• 3/4: correction replaces the stale current answer;
• 4/4: you can inspect the memory and explicitly confirm its deletion.

Reply with this short scorecard if you want to help improve the product:

AI tool:
Setup time:
Score out of 4:
One confusing step:

Please do not post credentials, private prompts, personal data, or confidential project details.

Free local or managed

• Free local: use the MCP setup and keep memory on your computer.
• Managed: Lians Personal is $10/month if you want a private hosted workspace, export and deletion controls, usage included, and email setup support.

You do not need to buy anything to run the challenge or share a result. The useful feedback is where the setup or memory behavior stops making sense.

[ebeirne]

Update: Lians Personal is live

The local and self-hosted Lians software stays free. If you want the same memory capability without running the service yourself, Lians Personal is now $10/month.

Personal includes:

• one managed account and private memory workspace;
• 100,000 memory writes and 50,000 recalls each month;
• controls to inspect, correct, delete, and export memories; and
• email help when setup gets confusing.

Lians does not change the underlying model or promise fewer tokens on every task. It helps when a small relevant recall can replace old conversation that would otherwise be sent again.

• Install Lians locally for free
• See Lians Personal

We are actively looking for everyday users—students, solo developers, researchers, and anyone who is tired of repeating context to Claude, Codex, Cursor, or another AI tool. If that is you, reply here with the tool you use and what you want it to remember.

[ebeirne]

Verified cross-agent result — August 14

We ran this challenge with synthetic data across two separate live AI clients:

• Cursor stored one project memory and a separate Cursor session recalled the exact codeword.
• A fresh Claude Code session authenticated to the same Lians account recalled the same exact codeword.
• After explicit confirmed deletion, fresh Cursor and Claude checks returned no matching memory.
• In a separate balanced Cursor large-rule workload, bounded context used 24.72% fewer provider-reported input tokens while preserving all four exact answers.

That token result is specific to the declared 201-line always-applied-rule workload; it is not a promise of universal savings or larger provider quotas. The reviewable methodology and aggregate evidence are in PR #163.

If cross-chat or cross-agent continuity is useful to you, run the challenge above and reply with your tool, setup time, score out of 4, and the most confusing step. If Lians earns a place in your workflow, star the repository so more users can find and pressure-test it.