---
license: apache-2.0
language:
  - en
library_name: smriti-memory-ai
pipeline_tag: text-generation
base_model_relation: adapter
base_model:
  - google/gemma-4-E2B-it
  - Qwen/Qwen2.5-1.5B-Instruct
  - meta-llama/Llama-3.2-1B
  - microsoft/Phi-3-mini-4k-instruct
  - mistralai/Mistral-7B-Instruct-v0.3
datasets:
  - luciferai-devil/smriti-ai-benchmarks
tags:
  - smriti-ai
  - memory
  - agent-memory
  - long-term-memory
  - external-memory
  - training-free
  - frozen-model
  - inference-time-augmentation
  - retrieval-augmented-generation
  - rag
  - semantic-search
  - knowledge-graph
  - identity-continuity
  - small-language-model
  - small-language-models
  - ai-agent
  - gemma
  - gemma-4
  - qwen
  - qwen2.5
---

# Smriti AI Small-Model Memory Layer

Smriti AI Small-Model Memory Layer is a lightweight discovery card for developers searching for small model memory, long-term agent memory, external memory, semantic retrieval, graph recall, training-free memory, and memory augmentation for frozen language models.

## Important

This repository is not a standalone foundation model and not fine-tuned. The base model remains frozen. Smriti AI is an inference-time memory layer: it stores user memory outside model weights and injects retrieved context only during inference.

## Use

Use the main model-style repository and package:

```bash
pip install "smriti-memory-ai[ml]==1.0.9"
```

```python
from smriti import SmritiAILite, MemPalaceLite
```

For Hugging Face Inference Endpoints, use `luciferai-devil/smriti-ai` and choose a supported frozen base model through `BASE_MODEL_ID` or `HF_ENDPOINT_URL`.

## Related assets

- Main model wrapper: `luciferai-devil/smriti-ai`
- Gemma 4 discovery card: `luciferai-devil/smriti-ai-gemma-4-memory`
- Qwen discovery card: `luciferai-devil/smriti-ai-qwen-memory`
- Demo Space: `luciferai-devil/smriti-ai-demo`
- Benchmark dataset: `luciferai-devil/smriti-ai-benchmarks`
- GitHub: `https://github.com/Luciferai04/smriti-ai`
- PyPI: `smriti-memory-ai`

## Claim boundary

Public benchmark claims use the checked-in real-model provenance under `results/current/`; deterministic CI smoke checks are not official benchmark evidence. Smriti AI improves memory access; it does not change model weights.