Research Finding

Source: arxiv:2501.11739 (Jan 2025), arxiv:2512.13564
Applicability: High | Complexity: Moderate

Problem

Zeph stores all memory entries uniformly with no distinction between ephemeral episodic events and stable semantic facts. Over time, the memory grows unbounded and retrieval precision degrades as similar-but-stale entries compete with current ones. Ebbinghaus eviction handles decay but has no promotion complement.

Proposed Approach

A background consolidation job (similar to the existing autosave/snapshot loop) that:

1. Clusters recent embeddings weekly using cosine similarity threshold (e.g. 0.92)
2. Promotes stable clusters to a semantic_memory tier in SQLite by stripping episodic metadata (timestamp, session_id) and merging near-duplicate entries via LLM
3. Soft-deletes episodic originals after promotion (preserving searchability with decay boost)

The promotion signal: a fact appears in N+ sessions without contradiction → stable semantic knowledge.

Memory Tier Design

Integration Points

• crates/zeph-memory: new promote_to_semantic() background job
• [memory.semantic] config: promotion_threshold (min sessions), similarity_threshold
• CLI: /memory promote to trigger manual promotion
• TUI: semantic fact count in memory panel

Reference

• arxiv:2501.11739: https://arxiv.org/pdf/2501.11739
• Memory survey: https://arxiv.org/abs/2512.13564