Continuum Memory Architectures for Long-Horizon LLM Agents#
Academic paper by Joe Logan (Mode7 GK, Tokyo) defining CMA as a formal architectural class for agent memory. Provides the theoretical foundation for why RAG is insufficient and what properties a real memory system must have.
The Problem with RAG#
RAG treats memory as a stateless lookup table: information persists indefinitely, retrieval is read-only, temporal continuity is absent. Agents reconstruct context afresh each time rather than maintaining it — no continuity of identity or purpose.
Six CMA Requirements (Necessary Conditions)#
- Persistence — state preserved across sessions. Agent accumulates identity.
- Selective Retention — memories compete for accessibility (recency, usage, salience). Ebbinghaus forgetting curves.
- Retrieval-Driven Mutation — every lookup alters future accessibility. Consulted fragments stabilize, contradictory ones recede.
- Associative Routing — structure connecting entities. Activation spreads along links for multi-hop answers.
- Temporal Continuity — episodic traces defined by order. Explicit temporal edges enable “what was happening around X?” queries.
- Consolidation and Abstraction — sleep-inspired replay and gist extraction. Episodes fade once schemas emerge.
Standard RAG meets none of these. These are necessary and collectively sufficient.
Reference Lifecycle#
Ingest → Retrieval → Mutation → Consolidation. Memory substrate is a structured store with semantic, temporal, and structural edges. Each node retains reinforcement history, salience, timestamps, provenance.
Evaluation Results (vs RAG Baseline)#
CMA won 82 of 92 decisive trials across four behavioral probes. Latency increased ~2.4× (1.48s vs 0.65s).
| Study | RAG Wins | CMA Wins | Effect Size |
|---|---|---|---|
| Knowledge Updates | 1 | 38 | d=1.84 |
| Temporal Association | 1 | 13 | h=2.06 |
| Associative Recall | 5 | 14 | h=0.99 |
| Disambiguation | 3 | 17 | h=1.55 |
Failure Modes#
- Latency scaling with graph edges
- Memory drift from reinforcement feedback loops
- Temporal sensitivity (half of temporal queries stumped both systems)
- Interpretability of evolving graphs
- Data governance for persistent memories
Key Insight#
Future agent architectures differentiated less by model size, more by how they instantiate memory. Memory is an inevitable architectural primitive.
Related Work#
mem0 (Mem0g), MemGPT, A-MEM (Zettelkasten, NeurIPS 2025), Hindsight (four-network architecture), SimpleMem, MemoRAG.