# AI use cases for Memoket Kite (memoket/memoket-kite)

Repo: [memoket/memoket-kite](https://github.com/memoket/memoket-kite) | Use cases: 3 | Page: https://dedupedai.com/use-cases/memoket-memoket-kite/

KITE by Memoket is an open-source, vector-free Python library that turns conversations into dated facts filed under topics and answers questions with the source lines behind them.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Answer who owns a client account now, with sources: https://dedupedai.com/use-cases/memoket-memoket-kite/answer-current-account-owner-with-sources/
- Stop assistants from guessing answers never stored: https://dedupedai.com/use-cases/memoket-memoket-kite/flag-unanswerable-questions-instead-of-guessing/
- Rerun the LoCoMo and LongMemEval benchmark scores: https://dedupedai.com/use-cases/memoket-memoket-kite/reproduce-locomo-and-longmemeval-scores/

## How to answer who owns a client account now, with sources with Memoket Kite

Repo: [memoket/memoket-kite](https://github.com/memoket/memoket-kite) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README

Store dated client messages and get the current answer with the exact lines it rests on.

**Who it is for:** An account manager or sales developer whose AI agent must say who is the decision-maker on an account as of today, after several changes over months.

**The problem:** Nearest-match memory returns an old statement that sounds relevant, such as the first owner, and gives no date or reason, so the agent gives a stale answer.

**The outcome:** The agent returns the current answer with each supporting moment, showing who said it and when, so a person can check it quickly.

### Steps

1. Install the library: `python -m pip install memoket-kite`
2. Set your key, replacing the placeholder with your own: `export OPENAI_API_KEY="sk-..."`
3. Create the artifacts folder: `mkdir -p artifacts`
4. Download the demo file: `curl -fsSL https://raw.githubusercontent.com/memoket/memoket-kite/main/examples/data/demo_codebook.xml -o artifacts/quickstart.xml` (inferred, check the README)
5. In Python, call Memory.load("artifacts/quickstart.xml"), then memory.remember([{"role": "user", "content": "<client message>"}], session_id="<session id>") for each dated client message.
6. Call memory.answer_with_evidence("Who is the decision-maker on <account> now?") and print result.text and each fact.sources.

### Prompt to paste into your agent

```text
Use the memoket-kite repo's Python API. Load artifacts/quickstart.xml, store each stored message about <client account> with remember, then call answer_with_evidence on "Who owns <client account> now?" Expected output: the current owner's name plus the source lines with their dates. Verify that the most recent dated source supports the answer before you report it.
```

### Caveat

The README says integrations with Claude Code, Codex, Cursor, and OpenCode are on the roadmap, so agents use KITE through the Python API today.

Risk, credentials: The setup puts your OpenAI API key in an environment variable, so keep it out of shared logs and scripts you publish.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/memoket-memoket-kite/answer-current-account-owner-with-sources/

## How to stop assistants from guessing answers never stored with Memoket Kite

Repo: [memoket/memoket-kite](https://github.com/memoket/memoket-kite) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README

Get an empty recall result for questions memory never covered, so the agent says it does not know.

**Who it is for:** A support or operations team lead whose chatbot answers questions about past customer interactions and must not invent details when nothing was recorded.

**The problem:** Most memory systems always return the closest match, so an assistant can confidently answer from unrelated text.

**The outcome:** When the memory has no matching fact, the recall comes back empty, and the agent can tell the user that the answer was not recorded.

### Steps

1. Install the library: `python -m pip install memoket-kite`
2. Set your key, replacing the placeholder with your own: `export OPENAI_API_KEY="sk-..."`
3. Create the artifacts folder: `mkdir -p artifacts`
4. Download the demo file: `curl -fsSL https://raw.githubusercontent.com/memoket/memoket-kite/main/examples/data/demo_codebook.xml -o artifacts/quickstart.xml` (inferred, check the README)
5. In Python, load the memory with Memory.load("artifacts/quickstart.xml"), then loop over memory.recall("<question never stored>") and count the facts returned.
6. If the list is empty, make the agent reply that the information was not recorded instead of producing an answer.

### Prompt to paste into your agent

```text
Use the memoket-kite repo's Python API. Load artifacts/quickstart.xml and run memory.recall on <question> that was never stored. Expected output: an empty result list, and the agent replies that the answer was not recorded. Verify the result list has zero items before the agent makes any claim.
```

### Caveat

The README describes the empty result behavior, but it does not say how a caller should word the reply to the user, so that wording is up to your agent.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/memoket-memoket-kite/flag-unanswerable-questions-instead-of-guessing/

## How to rerun the LoCoMo and LongMemEval benchmark scores with Memoket Kite

Repo: [memoket/memoket-kite](https://github.com/memoket/memoket-kite) | Difficulty: Advanced | Time: 2 hours | Basis: From the README

Rebuild the reported benchmark results locally to check KITE's scores before you adopt it.

**Who it is for:** A research engineer evaluating long-term memory options who needs to confirm published LoCoMo and LongMemEval-S scores under a shared protocol before choosing a memory engine.

**The problem:** Published benchmark numbers are hard to trust without a way to reproduce them under the same model, judge, and dataset revisions.

**The outcome:** You get a local run with manifests that record the commit, corpus digest, and scores, which you can verify offline.

### Steps

1. Read the benchmark guide at benchmarks/README.md in the repo.
2. Install the library: `python -m pip install memoket-kite`
3. Set your key, and if you use another provider also set OPENAI_BASE_URL, following the README.
4. Run the provided benchmark scripts named in benchmarks/README.md to rebuild the evaluation from the official datasets.
5. Check the local manifest and judging manifest, then run the verifier described in docs/reproducibility.md.

### Prompt to paste into your agent

```text
Use the memoket-kite repo. Follow benchmarks/README.md to run the LoCoMo evaluation with a configured compatible LLM endpoint. Expected output: a local judging manifest with an overall accuracy and an average reader context. Verify that the manifest's commit and corpus digest match the run before you compare the score with the README's 93.51%.
```

### Caveat

The README notes the benchmark corpora have their own licences in LICENSE-DATA.md, so check those terms before you redistribute any dataset files.

Risk, credentials: The benchmark run calls a paid LLM endpoint with your key, so keep the key private and watch the usage on your account.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/memoket-memoket-kite/reproduce-locomo-and-longmemeval-scores/