Conversational Memory

The transient, in-context buffer holding the immediate dialogue history. STM is implemented as a sliding window of recent messages serialized into the model's context window.

• Mechanism: Appends user and assistant turns to the prompt prefix.
• Constraint: Bounded by the model's maximum context length (e.g., 128k tokens).
• Failure Mode: Older turns are truncated, causing catastrophic context loss.
• Optimization: Uses conversation summarization to compress older turns into a dense summary before eviction.

A persistent, external knowledge store that survives individual sessions. LTM retains user preferences, facts, and entity relationships across conversations.

• Storage Backends: Implemented via vector databases (semantic recall) or knowledge graphs (structured facts).
• Retrieval: Queried dynamically using the current user intent as a search vector.
• Key Distinction: Unlike STM, LTM is session-independent and requires explicit write operations.
• Challenge: Preventing memory pollution where outdated or conflicting facts degrade response quality.

Records of specific past interactions stored as discrete event sequences. Episodic memory enables recall of 'what happened last time' rather than just extracted facts.

• Structure: Stores (timestamp, intent, entities, outcome) tuples.
• Use Case: 'Remember when we debugged the payment API last month?' triggers retrieval of that specific episode.
• Implementation: Often built on time-weighted vector retrieval where recency boosts relevance scores.
• Relationship: Bridges STM and LTM by promoting significant short-term episodes into long-term storage.

The active computational scratchpad holding intermediate reasoning results during a single turn. Working memory is distinct from dialogue history.

• Function: Stores chain-of-thought traces, tool call results, and partial computations.
• Lifetime: Purged after the final response token is generated.
• Architecture: Implemented via ReAct framework interleaving or explicit scratchpad tokens.
• Critical Role: Prevents the model from recomputing intermediate steps when executing multi-step tool chains.

Structured factual knowledge about entities and their relationships, independent of when the information was learned. Semantic memory answers 'what is X' rather than 'when did we discuss X'.

• Representation: Stored as subject-predicate-object triples in a knowledge graph.
• Query Pattern: 'What is the user's preferred database?' resolves against stored preferences.
• Update Mechanism: Facts are upserted based on explicit user statements or inferred from dialogue.
• Conflict Resolution: Requires truth maintenance to handle contradictory facts stated across sessions.

Implicit knowledge of how to execute multi-step workflows based on prior successful interaction patterns. Procedural memory encodes process rather than facts.

• Form: Stored as action sequences or workflow templates that can be replayed.
• Trigger: Activated when a user intent matches a known procedure signature.
• Example: A support agent remembering the 5-step diagnostic flow for a recurring error pattern.
• Implementation: Often encoded as few-shot examples in system prompts or retrieved workflow schemas from a template store.

The structured backbone of conversational memory. DST maintains a slot-value representation of user goals across turns. At each dialogue turn, the tracker updates a probability distribution over predefined slots (e.g., destination, date) based on the latest utterance and accumulated history. This explicit state representation enables deterministic business logic execution and prevents the model from forgetting constraints established earlier in the session.

The NLP task that binds pronouns and anaphoric expressions to their antecedents. Without coreference resolution, a system cannot understand that 'it' in 'Is it refundable?' refers to the 'flight booking' mentioned three turns ago. Modern systems use span-based neural models that score candidate mention clusters across the full dialogue history, linking that policy, the aforementioned, and hers to canonical entities stored in memory.

A compression technique that distills long dialogue histories into dense, lossy summaries. When a session exceeds the model's context window, a recursive summarization pipeline generates a running synopsis of key facts, decisions, and user preferences. This summary is injected into the system prompt as a memory prosthesis, trading granular recall for token efficiency. Critical for long-running agentic sessions.

Augments the user's current query with semantically relevant terms extracted from conversation history before retrieval. For example, if a user previously discussed 'Q4 budget constraints' and now asks 'show me options', the retriever receives an expanded query like 'show me options Q4 budget constraints affordable pricing'. This bridges the vocabulary gap between terse follow-up queries and the full informational need established across turns.

The cognitive capability to chain logical inferences across dialogue boundaries. Unlike single-turn Q&A, multi-turn reasoning requires the model to accumulate premises incrementally. A system must remember that Turn 1 established 'User is a premium subscriber' and Turn 3 asked about 'upgrade eligibility' to correctly infer in Turn 5 that the user qualifies for a loyalty discount without re-stating the facts.

A two-stage retrieval refinement where a cross-encoder model scores retrieved documents against the full dialogue context, not just the isolated query. The cross-encoder receives the concatenated history and each candidate document, outputting a relevance score that accounts for established constraints. This prevents the retriever from surfacing documents that match the current query lexically but contradict facts agreed upon earlier in the conversation.