Memory Types for Conversational AI Explained

Working Memory: The Active Conversation

Working memory in conversational AI corresponds to the current conversation context: the messages exchanged in the active session, the current topic, the user's immediate goals, and any temporary state like partially collected form data or active tool calls. In human cognition, working memory holds the information you are actively thinking about right now, with a famously limited capacity of roughly 7 items. In chatbot architecture, working memory is the context window, which similarly has a fixed capacity (measured in tokens rather than items) that limits how much information can be held simultaneously.

Implementation of working memory is straightforward: maintain an array of conversation messages and include them in each API call. The engineering challenges are capacity management (what to do when the conversation history exceeds the context window) and prioritization (which information stays in working memory when space is limited). Summarization techniques compress older turns into shorter representations, freeing capacity for new exchanges. Context window management strategies determine how much of the window is allocated to system prompt, conversation history, retrieved documents, and recalled memories.

Working memory is ephemeral by design. It serves the current conversation and is discarded when the session ends. Attempting to use working memory (raw conversation history) as a substitute for long-term memory produces poor results because raw conversation transcripts are verbose, full of low-value content (greetings, acknowledgments, filler), and organized by temporal sequence rather than by topic or relevance. Effective long-term memory requires extracting the important content from working memory and storing it in structured, searchable forms.

Episodic Memory: Past Interactions

Episodic memory stores records of specific past events and interactions. In human cognition, these are the memories of "what happened" with temporal and contextual anchoring: "I had coffee with Sarah last Tuesday and she mentioned the deadline moved to April." In conversational AI, episodic memories are records of specific past conversations: what was discussed, what decisions were made, what questions were asked, and what the outcome was.

Episodic memory enables a chatbot to reference specific past interactions: "Last time we talked about this, you mentioned that the API rate limit was causing problems. Did you find a solution?" This kind of reference creates a powerful sense of continuity because it demonstrates that the chatbot was not just present for the conversation but actually remembers the specific exchange. Users respond strongly to episodic recall because it mimics how human relationships work: we remember specific conversations with people, not just abstract facts about them.

Implementation of episodic memory involves storing conversation summaries or extracted events with temporal metadata (when the conversation happened), contextual metadata (what topics were discussed, which entities were mentioned), and outcome metadata (was the issue resolved, what was decided). Recall queries for episodic memory search by time range ("conversations from last week"), by topic ("conversations about API integration"), by entity ("conversations mentioning the staging server"), or by outcome ("unresolved issues"). Vector search alone is insufficient for episodic recall because it does not capture the temporal and contextual relationships that make episodic memories useful. A knowledge graph that connects conversations to topics, entities, and time periods through structured relationships provides much better episodic recall.

Semantic Memory: Accumulated Facts

Semantic memory stores facts, preferences, and knowledge that are detached from specific episodes. In human cognition, semantic memory is your knowledge of the world: "Paris is the capital of France," "I prefer window seats," "My sister's birthday is March 12." These facts were learned through specific experiences (episodes), but the knowledge itself is stored independently of when or how it was acquired. In conversational AI, semantic memory stores facts about the user and their world: their role, preferences, projects, technical stack, communication style, and domain knowledge.

Semantic memory is the most valuable type for conversational AI because it enables the chatbot to behave as though it knows the user as a person, not just as a sequence of past interactions. A chatbot with rich semantic memory about a user can: adapt its communication style to match the user's preferences (concise versus detailed, technical versus simplified), avoid asking questions it already knows the answers to (the user's name, role, company, plan), provide contextually relevant suggestions (recommending Python solutions because it knows the user works in Python), and personalize its responses based on the user's expertise level and interests.

Implementation of semantic memory requires extracting standalone facts from conversations and storing them with category labels, entity tags, and confidence scores. Unlike episodic memory (which is organized temporally), semantic memory is organized by topic and entity: all facts about the user's preferences, all facts about their current project, all facts about their technical environment. Recall queries for semantic memory search by relevance to the current conversation topic, with cognitive scoring that prioritizes recent, frequently accessed, and high-confidence facts. The key engineering challenge is keeping semantic memories accurate over time: facts change (users switch jobs, change preferences, update their tech stack), and the memory system must update stored facts when contradicting information arrives rather than accumulating conflicting entries.

Procedural Memory: Learned Patterns

Procedural memory stores how to do things rather than what things are. In human cognition, procedural memory covers skills and habits: how to ride a bicycle, how to type on a keyboard, how to navigate a familiar route. In conversational AI, procedural memory captures learned patterns of interaction: which response styles work best for this user, which troubleshooting steps typically resolve this type of issue, which topics this user frequently transitions to after discussing a given subject, and which information the user tends to need when asking a particular type of question.

Procedural memory is the least commonly implemented type in chatbot systems because it requires pattern recognition across multiple interactions rather than fact extraction from individual conversations. Building procedural memory involves analyzing conversation logs for recurring patterns: if users who ask about feature X always follow up with a question about feature Y, the chatbot can proactively mention Y when discussing X. If a specific user always asks for code examples after a conceptual explanation, the chatbot can include code examples preemptively for that user. These patterns emerge over dozens or hundreds of interactions, not from a single conversation.

Implementation of procedural memory requires statistical analysis of interaction patterns, either through explicit pattern mining (counting co-occurrence frequencies, tracking follow-up sequences) or through the memory system's consolidation process (which identifies patterns across stored memories and generates synthesized observations). Adaptive Recall's reflect tool performs this kind of cross-memory analysis, identifying patterns, connections, and insights that are not present in any individual memory but emerge from the collection as a whole.

How the Types Work Together

In practice, the four memory types operate simultaneously and inform each other. Working memory holds the active conversation and determines what to pay attention to right now. Episodic memory provides specific past interactions that are relevant to the current conversation. Semantic memory provides accumulated facts that shape the chatbot's understanding of the user. Procedural memory provides learned patterns that guide the chatbot's behavior and anticipate the user's needs.

A concrete example: a user says "same issue as before." Working memory has the current message. Episodic memory recalls the specific previous conversation where this issue was discussed (what the symptoms were, what troubleshooting was tried, what the resolution was). Semantic memory provides context about the user's environment (their tech stack, their infrastructure setup, their expertise level) that helps interpret the issue. Procedural memory notes that this user has reported the same issue three times in two months, suggesting a systemic cause rather than an isolated incident, and that previous resolutions involved restarting a specific service.

The chatbot's response draws on all four types: "This looks like the connection timeout issue we have discussed before. Last time, restarting the Redis cache resolved it, but since this is the third occurrence, it might be worth investigating the root cause. Based on your setup, the most likely culprit is the connection pool exhaustion we noticed in your configuration last month. Would you like to dig into that?" This response is only possible with all four memory types working together: episodic (specific past conversations), semantic (the user's configuration), procedural (pattern recognition across multiple incidents), and working (the current message and context).