# AI use cases for Voice Mem (xzf-thu/VoiceMem) Repo: [xzf-thu/VoiceMem](https://github.com/xzf-thu/VoiceMem) | Use cases: 5 | Page: https://dedupedai.com/use-cases/xzf-thu-voicemem/ VoiceMem is an open source memory system for real-time voice agents that stores facts and emotional context from speech and feeds relevant memories to a reply model. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Give a custom voice agent long-term memory: https://dedupedai.com/use-cases/xzf-thu-voicemem/add-long-term-memory-to-voice-agent/ - Store and search text facts with Voice Mem: https://dedupedai.com/use-cases/xzf-thu-voicemem/store-and-search-text-facts/ - Run the LoCoMo benchmark on a memory system: https://dedupedai.com/use-cases/xzf-thu-voicemem/benchmark-memory-on-locomo/ - Run the Voice Mem voice demo in a browser: https://dedupedai.com/use-cases/xzf-thu-voicemem/try-voicemem-voice-demo-locally/ - Fine-tune a Voice Mem model family adapter: https://dedupedai.com/use-cases/xzf-thu-voicemem/fine-tune-voicemem-adapter/ ## How to give a custom voice agent long-term memory with Voice Mem Repo: [xzf-thu/VoiceMem](https://github.com/xzf-thu/VoiceMem) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Plug Voice Mem into your own voice model so it recalls user facts and feelings across sessions. **Who it is for:** A developer building a real-time voice agent who wants it to remember a user's preferences and past conversations. **The problem:** A voice model forgets everything between sessions, so users repeat themselves and the agent cannot adapt to who they are. **The outcome:** Your reply model gets a short set of retrieved memories as context before it answers, and only the reply function changes. ### Steps 1. Clone the repo with `git clone https://github.com/xzf-thu/VoiceMem.git` and enter the folder with `cd VoiceMem`. 2. Install the memory system with `pip install voicemem`. 3. Download the default local models with `pip install -U huggingface_hub` and then `hf download zhifeixie/VoiceMem_Default_Models_Env --local-dir ./models`. 4. Set your key with `export OPENAI_API_KEY=sk-...`. 5. Run the sample agent with `python examples/03_simple_agent_with_voicemem_memory.py`, then replace the generation step with your own model in a function that takes text and memory_context. ### Prompt to paste into your agent ```text Using the VoiceMem repo at , wire my model into a reply function with signature my_reply(text, memory_context) and pass it to VoiceMem(reply=my_reply). Confirm the sample script runs and that memory_context is non-empty after I state a fact such as a food allergy. ``` ### Caveat Memory writes call the OpenAI API for fact extraction, while memory retrieval runs locally according to the README. Risk, credentials: The example needs an OPENAI_API_KEY, which should be kept in an environment variable and not committed to code. 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/xzf-thu-voicemem/add-long-term-memory-to-voice-agent/ ## How to store and search text facts with Voice Mem with Voice Mem Repo: [xzf-thu/VoiceMem](https://github.com/xzf-thu/VoiceMem) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README Save plain text facts in a fact-only mode and retrieve them with a search query. **Who it is for:** A developer who wants a simple memory store for user facts, without audio processing or emotion tracking. **The problem:** Building a per-user fact store with retrieval from scratch takes time, and the team wants a tested memory engine to start from. **The outcome:** A short Python script saves a fact such as a dietary restriction and returns it when asked a matching question. ### Steps 1. Clone the repo with `git clone https://github.com/xzf-thu/VoiceMem.git` and install with `pip install voicemem`. 2. Download the default models with `hf download zhifeixie/VoiceMem_Default_Models_Env --local-dir ./models` after running `pip install -U huggingface_hub`. 3. In a Python file, create Voice Mem(mode="leftbrain_only", openai_key=...) and call vm.ingest with a text fact. 4. Call vm.search with a question about that fact and print the result.result_leftbrain field. ### Prompt to paste into your agent ```text In the VoiceMem repo at , write a script that creates VoiceMem with mode leftbrain_only, ingests the text , and searches with the question . Expected output is the stored fact in result_leftbrain. Verify the search returns the fact after ingest. ``` ### Caveat The README describes this mode as storing factual text without emotion, so it does not replace the full dual-brain setup. 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/xzf-thu-voicemem/store-and-search-text-facts/ ## How to run the LoCoMo benchmark on a memory system with Voice Mem Repo: [xzf-thu/VoiceMem](https://github.com/xzf-thu/VoiceMem) | Difficulty: Advanced | Time: 1 hour | Basis: From the README Score Voice Mem on the LoCoMo dataset and check that the parser reads the data correctly first. **Who it is for:** A researcher or engineer who needs a reproducible score for a memory system and wants to compare it with published numbers. **The problem:** Memory benchmarks are easy to run wrong, for example by giving the reply model the full conversation instead of only retrieved memory. **The outcome:** A score report with accuracy by question type and median retrieval latency, produced by a documented command. ### Steps 1. Clone the repo and run `python evaluation/run.py --dataset locomo --data evaluation/examples/locomo_sample.json` as a small check. (inferred, check the README) 2. Run `python evaluation/run.py --dataset locomo --data data/locomo.json --inspect` to confirm the dataset parses without calling any model. (inferred, check the README) 3. Set `export OPENAI_API_KEY=sk-...` and run `python evaluation/run.py --dataset locomo --data data/locomo.json` for the full benchmark. (inferred, check the README) 4. Read evaluation/README.md for how to add another benchmark with one file and two functions. ### Prompt to paste into your agent ```text In the VoiceMem repo at , run the LoCoMo sample with evaluation/run.py, then run the full dataset only after the sample passes. Report the overall score, the per-category scores, and median retrieval latency. Verify that the answer model received retrieved memory and not raw conversation history. ``` ### Caveat The full run calls the OpenAI API, so the README's inspect mode should be used first to avoid unexpected charges. 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/xzf-thu-voicemem/benchmark-memory-on-locomo/ ## How to run the Voice Mem voice demo in a browser with Voice Mem Repo: [xzf-thu/VoiceMem](https://github.com/xzf-thu/VoiceMem) | Difficulty: Intermediate | Time: 45 minutes | Basis: From the README Start the local web demo and talk to a voice agent that remembers what you said. **Who it is for:** A developer or product person who wants to see the streaming voice memory working before building on it. **The problem:** Reading about a streaming memory architecture does not show how memory is retrieved before the speaker finishes a sentence. **The outcome:** A local web page at localhost:8787 where you speak, see transcripts, and check that a stated fact is recalled in a later turn. ### Steps 1. Clone the repo, install with `pip install voicemem`, and download models with `hf download zhifeixie/VoiceMem_Default_Models_Env --local-dir ./models`. 2. Run `python web/run.py` from the repo folder. 3. Open http://localhost:8787 in a browser and unmute the audio before playback. 4. Say a fact about yourself, then ask a question about it in a later turn and check the reply. ### Prompt to paste into your agent ```text In the VoiceMem repo at , start python web/run.py and open http://localhost:8787. Ask the user to state a fact, then to ask about it. Expected output is a reply that uses the stated fact. Verify the log file in results/logs was written. ``` ### Caveat The demo needs a microphone and browser audio permissions, and the README warns that playback needs audio unmuted. 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/xzf-thu-voicemem/try-voicemem-voice-demo-locally/ ## How to fine-tune a Voice Mem model family adapter with Voice Mem Repo: [xzf-thu/VoiceMem](https://github.com/xzf-thu/VoiceMem) | Difficulty: Advanced | Time: 3 hours | Basis: From the README Train your own adapter on the provided training script and reproduce the default configuration. **Who it is for:** A machine learning engineer who wants a custom Voice Mem memory-reading adapter for a different dataset or base model. **The problem:** Off-the-shelf models may not read Voice Mem memory well for a specific domain, so the team needs its own adapter. **The outcome:** A trained adapter checkpoint produced with the default settings, matching the configuration of the released checkpoint-3318. ### Steps 1. Clone the repo and install the training tools with `pip install ms-swift==4.5.2 bitsandbytes`. 2. Prepare training data in the format described in finetune/README.md. 3. Run `python finetune/train.py --data data/train.jsonl` with default parameters. 4. Check GPU memory requirements and base model options in finetune/README.md before changing the base model. ### Prompt to paste into your agent ```text In the VoiceMem repo at , install ms-swift==4.5.2 and bitsandbytes, then run finetune/train.py with my dataset . Expected output is a trained adapter. Verify the run uses the default parameters described in the README. ``` ### Caveat The README mentions GPU memory requirements without giving numbers here, so check finetune/README.md before starting a run. 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/xzf-thu-voicemem/fine-tune-voicemem-adapter/