Episodic, semantic, procedural memory; vector retrieval; memory management.

Long-term Memory — Persistent Agent Knowledge Stores

Long-term memory allows agents to retain knowledge across sessions and conversations. Three types: episodic (what happened — past interactions, outcomes), semantic (facts and knowledge — vector store retrieval), and procedural (how to do things — stored SOPs and learned workflows). Memory systems use vector databases for semantic retrieval, key-value stores for factual recall, and relational databases for structured episodic history. Memory management (updating, forgetting, consolidation) prevents unbounded growth.

Interactive Pipeline Flow

UnderstandImplementOptimise

🔍 Identify — what to remember

Not every message warrants long-term storage. Identify valuable memories: user preferences, important decisions, learned facts, recurring patterns, past mistakes.

Architecture Diagrams

Retrieving relevant memories and injecting into context.

Loading diagram...

Key Design Decisions

🧬Use mem0 or LangMem for managed memory

Building a custom memory system is complex. Libraries like mem0 handle embedding, deduplication, conflict resolution, and automatic extraction of memorable facts from conversations.

💉Inject memories into system prompt, not conversation

Add retrieved memories to the system prompt section: 'User preferences and context: [memories]'. Memories in the conversation history pollute the agent's dialogue flow.

🧹Memory without pruning becomes noise

After 6 months, a per-user memory store may have thousands of entries. Score memories by recency and retrieval frequency. Low-score memories dilute retrieval quality.

Discussion