research(memory): MemCollab cross-agent memory sharing — contrastive trajectory distillation (arXiv:2603.23234)

[bug-ops]

Source

arXiv:2603.23234 — MemCollab: Cross-Agent Memory Collaboration via Contrastive Trajectory Distillation (2026-03-24)

Technique

Multiple agents share a unified memory pool by:

1. Contrasting their reasoning trajectories on identical tasks to distill task-level patterns while filtering model-specific noise
2. Using category-aware retrieval at inference — memory entries are tagged with task category; recall filters by category before semantic search

Reports improved cross-agent knowledge transfer vs. naive shared memory.

Applicability to Zeph

Multi-agent orchestration: When Zeph's DagScheduler delegates subtasks across providers (e.g., planner on gpt-4o, executor on gpt-4o-mini), memory written during execution could be tagged with task category for cleaner cross-agent recall.

Category-aware retrieval maps onto the existing skill category grouping system ([data], [dev], [system] etc.) — extending that taxonomy to memory entries would require minimal changes to SemanticMemory.

MemCollab vs MAGMA: Complementary — MAGMA tracks typed entity relationships; MemCollab proposes task-category tags on episodic memory entries. Both could coexist.

Implementation sketch

• Add optional category field to MemoryEntry (in zeph-memory)
• Auto-tag entries based on active skill category or tool type during save
• Filter by category in SemanticMemory::search() when category context is available

Related

• research(memory): comprehensive memory survey — write-manage-read taxonomy, 5 mechanism families, open challenges (arXiv:2603.07670) #2351 — comprehensive memory survey
• research(memory): CraniMem gated episodic+knowledge-graph memory — goal-conditioned write gate, scheduled consolidation (arXiv:2603.15642) #2408 — CraniMem goal-conditioned write gate