Teaching Alfred to Remember with a Neuroscience-Inspired Memory System for AI Agents

Description and use

The author details the development of a neuroscience-inspired memory system for the AI assistant Alfred. By distinguishing between episodic and semantic memory, implementing temporal decay, and using a consolidation process, the agent gains the ability to recall past interactions, retain personal facts, and learn user preferences.

Problem

Personal AI agents often lack persistent memory, failing to recall past conversations, retain basic personal facts, or learn patterns from experience, resulting in an agent that feels like a 'capable stranger' rather than a personal assistant.

Outcome

The implementation of a neuroscience-inspired memory system transformed Alfred from a stateless task executor into a context-aware assistant capable of recalling past conversations, retaining personal facts, and learning user preferences through pattern recognition.

Workflow

• Adopt a cognitive science framework distinguishing between episodic memory (specific events) and semantic memory (distilled facts).
• Implement temporal decay using an exponential forgetting curve (e.g., Ebbinghaus's formula) to rank memories by recency and strength.
• Create tiered storage for episodic and semantic data with different lifespans.
• Build a memory consolidation pipeline ('dreaming') that runs in the background to promote important episodic patterns into semantic memory.
• Implement recall tracking to log memory retrieval frequency, boosting popular memories and fading unused ones.
• Use diversity ranking to suppress near-duplicate results in memory retrieval.
• Combine full-text search with vector similarity for hybrid search capabilities.
• Use content hashing to skip re-indexing unchanged content.
• Implement append-only memory writes to preserve history without overwriting.
• Provide chat tools (e.g., memory_search) allowing the agent to query its own memory during conversations.

Evidence

The formula looks like this: R(t) = e^(-t/S)... That is the exact formula Alfred uses to rank memories

Source

OpenClaw had already built something that went viral... Peter's approach to memory was almost aligned with what I had been reading in cognitive science.

Source