Implement LLM Conversation Memory Systems

This skill covers designing persistent memory systems for LLM conversations, spanning three memory types: short-term memory (recent interactions within a session), long-term memory (persistent across sessions), and entity memory (facts about people, places, and things). Storage and retrieval strategy design is covered; knowledge graph construction, semantic search implementation, and database administration are explicitly out of scope. The reference architecture is a tiered memory system with four layers: a buffer holding the current in-context conversation, short-term memory for recent session interactions, long-term memory persistent across sessions, and entity memory for facts extracted from messages. A consolidation process periodically moves short-term memories older than 24 hours into long-term storage when their importance score exceeds 0.7 or they were referenced more than twice, discarding the rest. Entity extraction uses an LLM call to pull structured entities and facts from each message, merging new facts into existing entity records while avoiding duplicate facts and tracking mention counts and timestamps. Memory-aware prompting assembles a prompt from known entities relevant to the query, semantically relevant long-term memories, and recent conversation history, then extracts any new memories from the model's response.

The skill also documents three severity-rated failure modes with fixes. Unbounded memory growth (HIGH severity) comes from storing every message with no cleanup or consolidation, causing slow retrieval and rising storage costs; the fix scores importance before storage and enforces a short-term memory limit that keeps the top 70% by importance and consolidates or discards the rest. Irrelevant retrieval (HIGH severity) comes from simple keyword matching with no relevance scoring, causing memories in context that don't help; the fix runs semantic search over a wider candidate set, scores each candidate's relevance to the query with an LLM call, and filters to results scoring above 0.5. Cross-user memory leakage (CRITICAL severity) comes from no user isolation in the memory store, letting one user see another's information; the fix namespaces every key by user ID, makes user ID a mandatory filter on every search, verifies ownership before delete, and adds GDPR-compliant per-user export and deletion.

Use it when building any conversational AI that needs to remember information across turns or sessions - conversation memory, remembering user facts, chat history, or long-term memory persistence.

Do NOT use it for knowledge graph construction, semantic search implementation, or database administration - those are explicitly out of scope. Delegate context-window or token limit problems to a context-window-management skill, retrieval or vector-search needs to a rag-implementation skill, and caching-strategy questions to a prompt-caching skill.

Requires knowledge of LLM conversation patterns, database basics, and key-value stores, plus familiarity with context-window-management and rag-implementation. Recommended primary tools: Mem0 (a memory layer for AI applications), LangChain Memory (memory utilities within LangChain), and Redis (an in-memory data store for session memory). Includes reference TypeScript interfaces and classes for the tiered memory system, entity extraction and upsert logic, importance-scored memory lifecycle management, relevance-filtered retrieval, and per-user isolated storage.

Works well with context-window-management, rag-implementation, prompt-caching, and llm-npc-dialogue. Delegates to context-window-management for context-window or token problems, to rag-implementation for retrieval or vector-search problems, and to prompt-caching for caching strategy. Composes with context-window-management and rag-implementation into a complete memory system: design memory tiers, implement storage and retrieval, integrate with context management, then add consolidation and cleanup.

Developers building conversational AI systems that need memory to persist within or across sessions, and who need concrete guidance on memory tiering, entity tracking, consolidation, and the sharp edges of unbounded growth, irrelevant retrieval, and cross-user data leakage.