Jepoxy/LFM2.5-350M-Memory-Extractor

LFM2.5-350M Memory Extractor

A lightweight fine-tune of LFM2.5-350M-Base for turning conversational text into concise, durable memories.

Model Description

This model was trained on 64K synthetic examples generated with a teacher model. For each example, the teacher was given a prompt and asked to extract the information that would be worth remembering over time.

The result is a small, specialized model focused on one task: distilling transient conversation into persistent memory.

Uses

It is intended for experimentation with memory extraction pipelines, especially in settings where running a larger model for every extraction would be unnecessarily expensive.

Note: This model is an experimental proof of concept, not a production-ready memory system. It was trained on only 64K synthetic samples, so output quality and consistency may vary significantly, especially outside the kinds of examples seen during training. The main goal of this release is to explore the idea of using a very small model for memory extraction. Think of it as a cool experiment and starting point rather than a finished product.

Example Format

Input: I love my cats

Output:
{
  "memories": [
    {
      "memory": "The user loves their cats",
      "confidence": "high"
    }
  ]
}

Direct Use

from transformers import pipeline

pipe = pipeline(
  "text-generation",
  model="Jepoxy/LFM2.5-350M-Memory-Extractor",
  device_map="auto",
)
  
messages = [
  {
  "role": "user",
  "content": "Recently moved to germany, what do you think about that?",
  }
]

output = pipe(
  messages,
  max_new_tokens=256,
  do_sample=False,
  clean_up_tokenization_spaces=False,
)

print(output[0]["generated_text"][-1]["content"]) #{"memories": [{"memory": "The user has recently moved to Germany.", "confidence": "high"}]}