Source

MemOS: A Memory OS for AI Systems (arxiv 2507.03724 / EMNLP 2025 oral)
Cross-attention memory retrieval with importance weighting (Frontiers 2025)

Core Idea

Assign an explicit importance score to each memory entry at write time, combining:

• Recency: how recently was this memory written
• Reference frequency: how many times has this memory been recalled
• Explicit salience: is the content emotionally/task-relevant (keyword heuristic or LLM score)

This "heat score" then weights the vector similarity score at retrieval time, so frequently-referenced or explicitly important memories surface higher than stale ones with similar embeddings.

Current Zeph Gap

Zeph uses temporal decay and MMR re-ranking, but lacks a persistent importance signal written at memory save time. Current retrieval ranking:

• Vector similarity (0.7 weight) + keyword match (0.3 weight)
• Temporal decay applied post-retrieval
• MMR lambda applied for diversity

What's missing: a reference count that increments each time a memory is recalled (positive signal), and an importance flag set at write time for high-salience content (e.g., user preferences, explicit instructions, key facts).

Implementation Sketch

Schema: add importance_score REAL DEFAULT 0.5 and recall_count INTEGER DEFAULT 0 columns to the memories table (additive SQLite migration, backward-compatible).

Increment recall_count each time a memory is returned by recall(). Compute importance at write time: combine content length (proxy for information density), explicit markers ("remember:", "important:", "always:"), and user feedback signal if available.

Blend into final retrieval score: final = alpha * vector_sim + beta * keyword + gamma * importance + decay

Complexity

Low-Medium. Schema migration is additive (new columns with defaults, no breaking changes). Scoring logic is pure Rust, no new dependencies. Retrieval formula tweak in recall_merge_and_rank() in zeph-memory.

Expected Benefit

• Higher-quality memory recall for frequently-referenced context
• User preferences and explicit instructions surface reliably
• Reduces retrieval noise from stale/irrelevant memories with coincidentally high embedding similarity

See Also

• SemanticMemory in zeph-memory/src/semantic.rs
• MMR re-ranking (already in place)
• Temporal decay (already in place -- this is complementary)
• MemOS: https://arxiv.org/pdf/2506.06326