Can a single model replace retrieval for long-term conversation memory?

The standard pipeline for long-term conversational memory is: (1) generate memories from past sessions, (2) store in a memory bank, (3) retrieve relevant memories via embedding similarity, (4) generate response using retrieved memories. COMEDY (Compressive Memory-Enhanced Dialogue Systems) collapses this into a single model that handles all four steps.

The departure is architectural: instead of storing discrete memory items and retrieving the most relevant ones, COMEDY reprocesses and condenses ALL past memories into a compressive representation with three dimensions:

1. Event recaps — concise summaries of what happened across all conversations, creating a historical narrative
2. User portraits — detailed user profile derived from conversational events
3. Relationship dynamics — how the user-chatbot relationship changes across sessions

This compressive memory inherently prioritizes salient information — unlike retrieval systems that must correctly rank relevance against a potentially vast database. The memory is always "up to date" because it is regenerated through compression, not queried from a static store.

Since Can long-context models resolve retriever-reader imbalance?, COMEDY takes this further: it eliminates the retriever entirely. The imbalance is resolved not by rebalancing, but by merging retrieval and generation into a single operation. The trade-off: compression necessarily loses some information, and there is no way to go back to the raw conversation for details that were compressed away.

The relationship dynamics dimension is particularly notable. Most memory systems track facts about the user (semantic memory) or events that occurred (episodic memory). Tracking how the relationship between user and agent evolves across sessions — increasing trust, shifting topic preferences, developing shared references — is a distinct memory type that neither retrieval nor summarization naturally captures.

Caveat from late-2025 work — COMEDY's One-for-All is exactly the consolidation pattern empirically shown to be fragile. Does agent memory degrade when continuously consolidated? demonstrates on controlled ARC-AGI Stream that an LLM continuously reprocessing past memories into compressed representations produces an inverted-U utility curve — memory utility rises, then degrades, then falls below no-memory baseline. COMEDY's architecture is the canonical instance of this pattern: a single model reprocesses and condenses ALL past memories on every update, with no retention of raw trajectories. The three failure mechanisms the empirical paper identifies — misgrouping experiences, stripping applicability conditions, narrow-stream overfitting — all apply to COMEDY's design by construction. The conversational setting differs from ARC-AGI in important ways (no objective ground truth, more forgiving evaluation, possibly more stable user preferences), so the regression may be less severe — but the architectural risk is real and was not visible at the time COMEDY was published. A design that also retained raw conversation snippets as first-class evidence — and gated the recompression step explicitly rather than firing it on every session — would inherit COMEDY's relationship-dynamics tracking while avoiding the consolidation-as-default failure mode. See also Why do LLM agents ignore condensed experience summaries? for the parallel evidence that agents systematically underuse condensed memory.