The Recall Memory Substrate offers a cutting-edge, local-first memory layer designed specifically for Large Language Model (LLM) agents. Unlike traditional memory setups that merely pull information on demand, Recall functions as an active memory substrate that intelligently manages the lifecycle of information, ensuring it remembers, self-corrects, and recalls data autonomously across sessions.

Key Features

• Autonomous Memory Management: Recall pushes updates rather than just pulling information. It operates in a continuous feedback loop with the agent, validating facts before actions and nesting new knowledge without user intervention.
• Local Storage: Eliminate dependencies on external cloud services. With SQLite as its core database, all memory is stored locally, ensuring privacy and access without needing an account.
• Efficient Memory Structure: Memory is represented in structured formats, allowing for easy retrieval and modification. Each write includes provenance, confidence levels, and rollback capabilities.

How It Works

Recall’s architecture revolves around a structured write schema, which ensures that every piece of information is submitted in a format that is validated by an admission firewall. This process includes:

1. Propose: The agent creates a memory proposal with all necessary details.
2. Admit: Each proposal is schema-validated to prevent incorrect or insecure entries.
3. Store: The processed memory is saved as addressable cells within the SQLite database, organized by relevance and evidence.
4. Compile: Contextual packets are generated that rank facts based on their effective confidence, allowing agents to access the most pertinent information based on their current tasks.

Usage Example

Integrate the Recall memory layer into any LLM agent with a simple command:

recall claude sync  # Set up for Claude Code
recall codex sync   # Set up for OpenAI Codex

After integration, agents will automatically consult Recall before executing tasks, maintaining accurate and updated context throughout their operations.

Demonstrations

Recall promotes truthfulness through a mechanism called supersession, where corrections do not overwrite previous data but create new records that challenge outdated information. For example:

• When a value needs to be updated, a new entry is created that specifies it contradicts the older version.
• The system demotes the previous entry rather than deleting it, providing a clear audit trail that maintains the integrity of past decisions.

An engaging demo showcases the supersession functionality:

This ensures that no information is lost while allowing for dynamic and responsive updates to the knowledge base.

Conclusion

The Recall Memory Substrate redefines memory management for LLM agents by providing a reliable, local-first approach that fosters integrity, accountability, and intelligent action based on historical context. This level of control and transparency is unique among memory systems, making it an invaluable tool for developers looking to enhance the capabilities of their AI agents.

Blog post explaining what Push vs Pull memory actually is: https://dev.to/hendrixxcnc/push-vs-pull-memory-a-better-way-to-think-about-ai-agent-memory-3lnp