permem 0.1.0

Persistent memory for AI - add memory to any LLM in one line

Permem Python SDK

Persistent memory for AI - add memory to any LLM in one line.

Installation

pip install permem

Quick Start

import asyncio
import permem

async def main():
    # Store a memory
    await permem.memorize("User's name is Ashish")

    # Recall memories
    result = await permem.recall("What is the user's name?")
    for memory in result.memories:
        print(memory["summary"])

asyncio.run(main())

Usage Modes

Tools Mode (Manual Control)

from permem import Permem

async def main():
    mem = Permem(user_id="user-123")

    # Store memories
    await mem.memorize("User prefers dark mode")

    # Search memories
    result = await mem.recall("user preferences", limit=5)

Auto Mode (Automatic Memory Management)

from permem import Permem

async def chat_with_memory(user_message: str, user_id: str):
    async with Permem(api_key="pk_your_api_key") as mem:
        # Before LLM call - inject relevant memories
        context = await mem.inject(user_message, user_id=user_id)

        system_prompt = "You are a helpful assistant."
        if context.should_inject:
            system_prompt += f"\n\n{context.injection_text}"

        # ... call your LLM ...
        assistant_response = "Nice to meet you!"

        # After LLM response - extract new memories
        messages = [
            {"role": "user", "content": user_message},
            {"role": "assistant", "content": assistant_response}
        ]
        await mem.extract(messages, user_id=user_id)

Configuration

Environment Variables

export PERMEM_URL="http://localhost:3333"
export PERMEM_USER_ID="default"
export PERMEM_API_KEY="your-api-key"

Programmatic Configuration

from permem import Permem, configure

# Configure singleton
configure(
    url="http://localhost:3333",
    user_id="user-123",
    api_key="your-api-key"
)

# Or create instance
mem = Permem(
    url="http://localhost:3333",
    user_id="user-123",
    api_key="your-api-key",
    max_context_length=8000
)

API Reference

memorize(content, conversation_id=None, async_mode=False)

Store a memory.

recall(query, limit=5, mode="balanced", conversation_id=None)

Search memories by semantic similarity.

Modes:

• "focused" - Higher precision, fewer results
• "balanced" - Default balance
• "creative" - Broader matches

inject(message, user_id, context_length=0, conversation_id=None)

Get relevant memories before LLM call. Returns InjectResponse with memories, injection_text, and should_inject.

extract(messages, user_id, context_length=None, conversation_id=None, extract_threshold=None, async_mode=False)

Extract memories from conversation after LLM response. Returns ExtractResponse with should_extract, extracted, and skipped_duplicates.

health()

Check if the Permem server is healthy.

License

MIT