ADR-0006: Episodic Conversation Architecture¶

1. Context¶

AI agents engaged in continuous interaction require conversational memory that captures multi-turn dialogues, user intents, temporal sequences, and contextual nuance. Flat vector databases store disjoint text fragments, losing the chronological sequence and multi-speaker turn structure inherent to human conversation. Spector's episodic memory layer must model conversation as a first-class cognitive entity.

2. Problem Statement¶

Traditional naive RAG approaches suffer from three severe conversational memory failures:

1. Loss of temporal continuity: Retrieving individual dialogue turns via cosine similarity loses conversational context and turn order.
2. Speaker attribution collapse: Inability to differentiate between user inputs, agent responses, tool execution results, and environmental observations.
3. Context window saturation: Storing raw verbatim transcripts rapidly exceeds LLM context windows, requiring intelligent summarization and episodic clustering.

3. Decision Drivers¶

• Turn-Preserving Temporal Chains: Episodic records must maintain causal and chronological links to preceding and succeeding turns.
• Multi-Role Attribution: Clear off-heap structural representation of speaker identities (User, Agent, Tool, System).
• Episodic Boundary Detection: Automatic segmentation of continuous conversational streams into coherent episodic sessions.
• Hybrid Retrieval Integration: Fused recall combining chronological recency, emotional valence, and semantic similarity.

4. Considered Options¶

Option 1: Document-Per-Conversation Chunking¶

• Description: Group conversations into fixed-token documents (e.g. 500 tokens) and embed each chunk.
• Advantages: Simple integration with conventional vector stores.
• Disadvantages: Splits conversational turns arbitrarily; loses fine-grained speaker attribution and temporal trajectory.

Option 2: External Knowledge Graph Storage (Neo4j)¶

• Description: Store conversations as nodes and edges in an external graph database.
• Advantages: Expressive graph query capabilities (Cypher).
• Disadvantages: Heavy external infrastructure dependency; introduces network query hops incompatible with sub-millisecond SLAs.

Option 3: Off-Heap Episodic Threading & Temporal Hyperedges (Selected)¶

• Description: Represent each conversational turn as an off-heap episodic engram with explicit speaker metadata, timestamp, and forward/backward pointers (prevEpisodeId, nextEpisodeId). Group conversational sessions via temporal hyperedges in HyperEntityGraphMemory, enabling multi-turn context expansion during recall.
• Advantages: Zero-GC off-heap layout, sub-millisecond sequential traversal, exact speaker attribution, and seamless integration with the cognitive scoring pipeline.
• Disadvantages: Requires custom off-heap indexing for session-based conversational lookups.

5. Decision Outcome¶

Chosen Option: Option 3 (Off-Heap Episodic Threading & Temporal Hyperedges).

Positive Consequences¶

• Retains complete conversational trajectories across hundreds of turns.
• Traversal APIs allow the LLM to reconstruct the surrounding context window of any retrieved memory turn.
• Integrates directly with ReflectDaemon to consolidate episodic conversation threads into semantic facts during idle sleep cycles.

Negative Consequences & Trade-offs¶

• Additional metadata overhead per conversational turn (pointers, role ordinals, timestamps).
• Session boundaries must be detected using time-gap heuristics or explicit conversation delimiters.

6. Pros and Cons of the Options¶

7. Implementation Plan¶

1. Phase 1: Define EpisodicTurnRecord layout in spector-kernel with role flags and pointer fields.
2. Phase 2: Implement conversational ingestion pipelines in spector-ingestion that link sequential turns.
3. Phase 3: Add temporal context expansion to RecallPipeline (retrieving \(K\) turns before and after high-scoring matches).
4. Phase 4: Implement automated conversation session boundary detection based on inactivity thresholds.

8. Code Reference & Verification¶

• Primary Module(s): memory/spector-memory, nucleus/spector-core
• Key Packages: com.spectrayan.spector.memory.episodic, com.spectrayan.spector.memory.pipeline
• Classes: EpisodicMemoryStore.java, ConversationSessionManager.java, TemporalSequenceRelay.java
• Verification Tests: EpisodicConversationIntegrationTest.java, TemporalContextExpansionTest.java