Description

arXiv:2604.01007 (April 1, 2026) — Omni-SimpleMem: Autoresearch-Guided Discovery of Lifelong Multimodal Agent Memory by Liu et al. (UNC-Chapel Hill, UC Berkeley, UC Santa Cruz, Cisco).

OmniMem proposes an autonomous research pipeline that systematically discovers and validates memory architecture improvements through self-experimentation — running ~50 experiments across two benchmarks to diagnose failure modes and propose architectural modifications.

Code: https://github.com/aiming-lab/OmniMem

Key Results

• LoCoMo benchmark: F₁ improved from 0.117 → 0.598 (+411%)
• Mem-Gallery benchmark: F₁ improved from 0.254 → 0.797 (+214%)

What the Autoresearch Pipeline Discovered

Top gains by category:

• Bug fixes: +175% — the autoresearch pipeline finds and patches its own memory bugs
• Prompt engineering: +188% on specific categories
• Architectural changes: +44%

Crucially, architecture changes from self-experimentation outperformed all hyperparameter tuning combined.

Relevance to Zeph

Zeph's self-learning system (zeph-skills, ReasoningMemory, SkilHeuristics) already accumulates outcomes and extracts heuristics. OmniMem extends this concept to the memory architecture itself:

1. Memory architecture self-improvement: Rather than only tuning skill strategies, the pipeline diagnoses why memory retrieval fails and proposes architectural patches
2. Multimodal memory: OmniMem handles image + text memories; Zeph is currently text-only. This opens a path to image/attachment memory.
3. Autoresearch evaluation loop: The pipeline runs automated evaluations (LoCoMo-style) without human oversight — a natural fit for Zeph's CI cycle

Proposed Design Direction

• Extend the skill_outcomes / reasoning_strategies self-learning loop to also log memory retrieval failures (no-hit turns, low-confidence recalls)
• Add a periodic background task (via zeph-scheduler) that runs a self-evaluation micro-benchmark on recent memory retrievals
• Use retrieval failure analysis to tune SYNAPSE spreading activation parameters (decay factor, depth, MMR threshold)
• Track memory architecture improvement suggestions in skill_heuristics table, distinct from skill heuristics

References

• Paper: https://arxiv.org/abs/2604.01007
• Code: https://github.com/aiming-lab/OmniMem
• Related: research(memory): ReasoningBank — distill generalizable reasoning strategies from agent successes/failures (arXiv:2509.25140) #3312 (ReasoningBank), research(memory): MemCoT test-time memory scaling — multi-view long-context reasoning with episodic + semantic short-term memory #3564 (MemCoT), research(memory): MemReader active memory extraction — teach agent when NOT to write to long-term memory #3222 (MemReader), APEX-MEM spec