# AI use cases for MemT 1 0 (MemTensor/MemT-1.0)

Repo: [MemTensor/MemT-1.0](https://github.com/MemTensor/MemT-1.0) | Use cases: 3 | Page: https://dedupedai.com/use-cases/memtensor-memt-1-0/

MemT-1.0 is an official research repository for a memory foundation model that keeps persistent, evolving memory inside the model's computation, served through SGLang with a Python client.

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

- Run a chat that remembers facts across turns: https://dedupedai.com/use-cases/memtensor-memt-1-0/run-persistent-memory-chat-locally/
- Save a memory snapshot and resume it in a new session: https://dedupedai.com/use-cases/memtensor-memt-1-0/save-and-restore-memory-snapshot/
- Compare replies with native memory on and off: https://dedupedai.com/use-cases/memtensor-memt-1-0/compare-memory-on-and-off-replies/

## How to run a chat that remembers facts across turns with MemT 1 0

Repo: [MemTensor/MemT-1.0](https://github.com/MemTensor/MemT-1.0) | Difficulty: Advanced | Time: 2 hours | Basis: From the README

Serve MemT-1.0-Nano with SGLang and chat with native memory that is kept after each reply.

**Who it is for:** An ML engineer with a Linux GPU server who wants to test whether a model can keep user facts in its own memory instead of a long transcript.

**The problem:** Long chat transcripts grow costly, and it is hard to tell whether a model retains earlier facts without resending the full context.

**The outcome:** A running SGLang server and a Python session where facts from earlier turns, such as a pet's name and age, are answered correctly after the plaintext context is cleared.

### Steps

1. Run `cd inference` in a clone of the MemTensor/MemT-1.0 repository.
2. Run `uv pip install '.[server]'` on the Linux GPU server.
3. Run `sh start_memt_nano.sh /models/MemT-1.0-Nano` and keep it running; it serves at http://127.0.0.1:30000.
4. In another terminal, run the Python connection code from the Quick Start, pointing model_path at the same checkpoint.
5. Call `session.chat(..., enable_memory=True)` with a fact in one turn, then ask a follow-up question that depends on that fact. (inferred, check the README)

### Prompt to paste into your agent

```text
Use the MemTensor/MemT-1.0 repo to start the MemT-1.0-Nano server from /models/<checkpoint-folder> and run the Quick Start chat example. Expected output: a reply that restates the kitten name, age and favorite food from earlier turns. Verify that the answer names facts that were given only in the first message.
```

### Caveat

The README requires a complete local checkpoint and a CUDA GPU setup, and it notes that the sample replies are illustrative rather than guaranteed output.

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/memtensor-memt-1-0/run-persistent-memory-chat-locally/

## How to save a memory snapshot and resume it in a new session with MemT 1 0

Repo: [MemTensor/MemT-1.0](https://github.com/MemTensor/MemT-1.0) | Difficulty: Advanced | Time: 2 hours | Basis: From the README

Store a model's memory state to a file and restore it later without the old chat transcript.

**Who it is for:** A developer building an assistant that must pick up where a user left off after a restart or a new process.

**The problem:** Restarting a conversation normally loses everything the model learned about the user, or forces the app to store and resend full transcripts.

**The outcome:** A snapshot file in the server's state directory that a new session can load, with zero stored messages and the earlier facts still answerable.

### Steps

1. Complete the server setup from the persistent memory chat use case so `start_memt_nano.sh` is running.
2. In your Python session, call `snapshot = session.save_state("memmy_memory.safetensors")` to commit pending turns and save memory.
3. Run `session.close()` to end the original session.
4. Start a new `MemTSession` with `state=snapshot["path"]` and check that `len(restored.messages)` prints 0.
5. Ask a question that depends on the saved facts and confirm the reply uses them.

### Prompt to paste into your agent

```text
Using the MemTensor/MemT-1.0 repo and a running MemT-1.0-Nano server, save memory from a session to <snapshot-name>.safetensors, then restore it in a new MemTSession. Expected output: restored.messages has length 0 and a reply to a fact-based question is correct. Verify the returned snapshot path is used as the state argument.
```

### Caveat

Snapshots are saved in the server's configured state directory, so the restored session only works against a server that can read that directory.

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/memtensor-memt-1-0/save-and-restore-memory-snapshot/

## How to compare replies with native memory on and off with MemT 1 0

Repo: [MemTensor/MemT-1.0](https://github.com/MemTensor/MemT-1.0) | Difficulty: Intermediate | Time: 1 hour 30 minutes | Basis: From the README

Run the same question with memory reading disabled to see what the memory layer changes.

**Who it is for:** A researcher or evaluator testing a memory model who needs a plain baseline next to memory-enabled answers.

**The problem:** Without a baseline, it is hard to tell whether a good answer came from native memory or from the model's general knowledge.

**The outcome:** Two answers to the same prompt, one with memory reading on and one with it off, that can be compared side by side.

### Steps

1. Start the Nano server with `sh start_memt_nano.sh /models/MemT-1.0-Nano` on the GPU host.
2. Create a session with `MemTSession(client, chat_format, parser)` and store a fact using `enable_memory=True`.
3. In a fresh session, call `chat.set_memory_read(False)` and ask the same fact-based question with `enable_memory=False`.
4. Repeat the question in a session where memory reading is on, and record both answers.

### Prompt to paste into your agent

```text
Use the MemTensor/MemT-1.0 repo with a running MemT-1.0-Nano server to ask <question> with memory reading on and off. Expected output: two answers, where only the memory-on answer uses facts from the saved conversation. Verify that the memory-off session was created without restoring a snapshot.
```

### Caveat

The README reports benchmark results for 11 memory benchmarks, but this repo's examples are small illustrations and do not reproduce those evaluations.

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/memtensor-memt-1-0/compare-memory-on-and-off-replies/