Angel Recall

Angel Recall is a Python implementation of the MemOS (Memory Operating System) architecture, designed to give AI agents a long-term, evolvable memory. It uses a Lane-Based Command Queue instead of concurrent async to ensure predictable, race-free execution.

The Lane-Based Advantage

Unlike standard async libraries, Angel Recall processes tasks for each session (user) in a dedicated Lane.

• Predictable Execution: Tasks within a session run sequentially, preventing race conditions.
• Clean Logs: No interleaved garbage in your application logs.
• Simplified Mental Model: You don't need to worry about asyncio locks or complex parallel debugging.

Quick Start

Installation

pip install angel-recall

Memory Dashboard

Angel Recall comes with a built-in Flask dashboard to visualize your memory vault and chat with your agent in real-time.

# Launch the dashboard
angel-recall memos

The dashboard allows you to:

• Configure API Keys: Add OPENAI_API_KEY, ANTHROPIC_API_KEY, etc., directly from the UI.
• Switch Models: Test different LLMs (local via Ollama or remote via LiteLLM).
• Inspect Memories: View all MemCubes in your vault, including their semantic types, access scopes, and timestamps.
• Chat: Interact with the memory-augmented agent and see how memories are created and retrieved.

Basic Usage

from angel_recall import MemOS

# Initialize the Memory OS (Synchronous)
memos = MemOS(persist_directory="./my_vault")

# Store a new memory
memos.process("Remember that I prefer technical deep-dives.")

# Retrieve relevant memories
response = memos.process("How should you format the report for me?")
print(response["response"])

Configuration

Angel Recall is designed to be flexible. You can configure the model and persistence directory during initialization:

memos = MemOS(
    persist_directory="./my_vault",
    model="ollama/gemma3n:e4b" # Supports any LiteLLM-compatible model
)

Local Embedding Optimization

By default, angel-recall relies on external calls for embeddings via litellm. For better performance, privacy, and zero-latency retrieval, you can enable local embeddings using sentence-transformers.

Enabling Local Embeddings

You can enable local embeddings during initialization or dynamically:

# During initialization
memos = MemOS(local_embedding=True)

# Or dynamically
memos.enable_local_embedding(True)

Installation

To use this feature, you must manually install the sentence-transformers library:

pip install sentence-transformers

If local_embedding is set to True but the library is not found, angel-recall will print an error message and exit.

Intelligent Memory Categorization

Angel Recall goes beyond basic storage by automatically categorizing and prioritizing your memories based on their content and intent:

• Corrections: When a user says "Actually, it's X," the system identifies this as a CORRECTION.
  • High Priority: Corrections are assigned a higher priority (default 5) to ensure they override older information in retrieval.
  • Proactive Archiving: Storing a correction automatically archives older conflicting facts to maintain context accuracy.
• Insights: When the agent or user identifies a complex pattern or deep understanding, it is stored as an INSIGHT (priority 3).
• Facts & Preferences: Standard user data (priority 0).

Memory Policies

Angel Recall implements several policies to ensure your agent's memory stays relevant and secure:

• Weight Decay & Forgetting: To simulate how humans prioritize information, memories have a weight (default 1.0). In every processing cycle, weights decay by a configurable rate.
  • Reinforcement: Accessing or "touching" a memory increases its weight.
  • Forgetting: If a memory's weight falls below a threshold (e.g., 0.3), it is automatically moved to MemoryState.ARCHIVED.
  • Configuration:

    memos = MemOS(decay_rate=0.01, min_weight=0.5)
    

• Lifecycle Management: Memories automatically transition through states (GENERATED -> ACTIVATED -> ARCHIVED). Cold memories are eventually moved to long-term storage to keep the retrieval context clean.
• TTL (Time-to-Live): You can set an expiration for any memory. Once the TTL is reached, the memory is automatically purged during the next process cycle.
• Access Governance: Supports PRIVATE, SHARED, and PUBLIC scopes. The system verifies ownership and permissions before any read or write operation.
• Sensitivity Masking: Built-in support for PII redaction and sensitive tag handling to prevent leakage of private data into LLM prompts.

Access Governance Example

You can control the visibility of memories programmatically or via tools. By default, all memories are private to the owner.

from angel_recall import MemOS, AccessScope

memos = MemOS()

# Programmatically update access
# Only the owner ('alice') can change the scope
success = memos.api.update_access_scope(
    cube_id="some-uuid",
    scope=AccessScope.SHARED,
    user="alice"
)

if success:
    print("Memory is now shared with other users in the system.")

When using get_memory_tools, your agent can also manage its own privacy:

# The agent can call 'set_memory_access' tool
# Example: set_memory_access(memory_id="...", scope="public")

Advanced Usage: Agent Tools

If you are building your own LangGraph agents and want to give them explicit control over memory, you can use get_memory_tools. This is perfect for complex agents that need to decide when to commit something to long-term storage.

from langgraph.prebuilt import ToolNode
from angel_recall import MemOS, get_memory_tools

memos = MemOS()
tools = get_memory_tools(memos, user="alice")

# Create a tool node for your graph
tool_node = ToolNode(tools)

# Bind tools to your model
# model_with_tools = model.bind_tools(tools)

Tool-based LangGraph Example

from langgraph.graph import StateGraph, END
from angel_recall import MemOS, get_memory_tools

memos = MemOS()
tools = get_memory_tools(memos, user="alice")

def call_model(state):
    # Your logic here to call an LLM bound with 'tools'
    pass

workflow = StateGraph(MessagesState)
workflow.add_node("agent", call_model)
workflow.add_node("tools", ToolNode(tools))
workflow.set_entry_point("agent")
# Add conditional edges for tool usage...

Testing

We use pytest for testing. To run the tests using the qwen2.5:0.5b model:

pytest tests/test_core.py

Ensure Ollama is running and you have pulled the model: ollama pull qwen2.5:0.5b.

Core Components

• MemVault: The storage engine combining ChromaDB for semantic search and NetworkX for relationship mapping.
• MemReader: An intelligent parser that understands intent (store, retrieve, delete) and categorizes memories into semantic types like CORRECTION, INSIGHT, FACT, or PREFERENCE.
• MemGovernance: Ensures privacy and access control, managing who can read or modify specific memory cubes.
• MemScheduler: Dynamically selects the best memory fragments based on the current task's context.

Contributing

We welcome contributions that improve the efficiency of the memory scheduler or add support for new vector backends. Please see our contributing guidelines for more details.

License

Apache License 2.0. See LICENSE for details.

Credits

• MemOs paper
• Memelord by Glauber Costa