Description

HeLa-Mem (Hebbian Learning and Associative Memory) introduces a bio-inspired dual-layer memory architecture for LLM agents based on cognitive neuroscience mechanisms. The paper addresses coherence across extended interactions that go beyond fixed context windows.

Key Approach

• Episodic Memory Graph: evolves via Hebbian learning — repeated co-activation strengthens connections between related memories
• Hebbian Distillation: a Reflective Agent identifies densely connected memory hubs in the episodic graph and promotes them into semantic (reusable) knowledge
• Three mechanisms: association (co-activation strengthening), consolidation (episodic→semantic), spreading activation (dynamic propagation)

Results

• Improved performance on LoCoMo benchmark
• Reduced token consumption compared to flat embedding stores
• Captures how knowledge organizes (not just what is stored) via episodic-semantic distinction

Relevance to Zeph

Zeph's currently uses semantic similarity retrieval on flat embedding vectors. HeLa-Mem's Hebbian Distillation pattern could improve the experience store by:

• Upgrading episodic edge weights based on access frequency (Hebbian reinforcement)
• Automatically promoting hot clusters to the semantic/skills layer
• Reducing retrieval cost through structured activation propagation instead of brute-force ANN search

Implementation Sketch

1. Add edge weight field to graph (already in zeph-memory as petgraph)
2. On each memory access: increment edge weight between co-activated nodes
3. Periodic consolidation pass: identify high-degree/high-weight clusters → extract as skills or persistent rules
4. Retrieval: spreading activation from query node vs. current cosine-only ANN

Source

• Paper: https://arxiv.org/abs/2604.16839
• Benchmark: LoCoMo (long conversation memory)

Spec: not yet created — assess in next research cycle.