Description

arXiv:2604.08216 (April 9, 2026) — MemCoT: Test-Time Scaling through Memory-Driven Chain-of-Thought.

MemCoT is a plug-and-play, training-free framework that addresses LLM hallucination and catastrophic forgetting during causal reasoning over massive, fragmented long contexts. Existing memory mechanisms treat retrieval as a static, single-step passive matching process, leading to semantic dilution and contextual fragmentation.

Core Components

Multi-view Long-Term Memory Perception Module:

• Zoom-In: evidence localization — identifies where relevant evidence resides in the long context
• Zoom-Out: contextual expansion — reconstructs surrounding causal structure necessary for reasoning

Task-Conditioned Dual Short-Term Memory:

• Semantic state memory: tracks conceptual state across reasoning steps
• Episodic trajectory memory: tracks the reasoning path itself (what was inferred, in what order)

Results

On the LoCoMo benchmark (long-context conversation memory), MemCoT achieves SOTA:

• GPT-4o-mini with MemCoT: F₁ = 58.03 (vs baseline ~30)
• Qwen2.5-14B with MemCoT: F₁ = 57.06

Compact models outperform much larger baselines on long-context memory tasks.

Relevance to Zeph

Zeph's memory system (MAGMA graph + SYNAPSE spreading activation + scene consolidation) is strong at structural/associative retrieval but doesn't address the reasoning dimension over long contexts. MemCoT's dual short-term memory maps naturally to Zeph's context builder:

• Semantic state memory ↔ extend TurnContext with a mutable "semantic state" accumulator updated after each LLM turn
• Episodic trajectory memory ↔ the reasoning-bank in ReasoningMemory already stores trajectories; MemCoT's Zoom-In/Out could feed the retrieval step
• Multi-view LTM perception ↔ MemoryOrchestrator::recall could be extended with a zoom-in pass (localize) before the zoom-out pass (expand context)

This is training-free — it operates entirely at inference time through prompt engineering + retrieval structuring.

Proposed Design Direction

• Add a SemanticStateAccumulator component to TurnContext that summarizes the "conceptual state" after each turn
• Extend ReasoningMemory retrieval with Zoom-In (targeted evidence localization) before the existing spreading activation
• Expose a memory.memcot.enabled config flag; default off; benchmark improvement on multi-turn factual tasks

References

• Paper: https://arxiv.org/abs/2604.08216
• Related: research(memory): ReasoningBank — distill generalizable reasoning strategies from agent successes/failures (arXiv:2509.25140) #3312 (ReasoningBank), research(core): Focus agent — model-controlled active context compression with Knowledge block (arXiv:2601.07190) #3313 (Focus agent), APEX-MEM spec /specs/004-memory/004-7-memory-apex-magma.md