SORNMemory: A Spiking Neural Network-Driven Episodic Memory System for LLMs

SORNMemory is an open-source project maintained by omgbox on GitHub, released in June 2025 with the original title "SORNMemory: SORN-powered episodic memory for LLMs". It leverages Self-Organizing Recurrent Neural Networks (SORN) to provide large language models (LLMs) with long-term episodic memory beyond context window limits, using biologically inspired mechanisms instead of traditional methods like vector databases or sliding windows.

Current LLMs face hard context window limits (8K-128K tokens). Existing solutions have flaws: sliding window truncates early information, vector databases rely on external embedding models, and summary compression loses details. SORNMemory proposes storing dialogue history as spike-weight patterns instead of raw text or vectors to address these issues.

The SORN network has 375 neurons and ~13,000 synapses (runs on CPU) and uses 5 plasticity rules: STDP (spike-timing-dependent plasticity), ISTD (inhibitory STDP), synaptic scaling, intrinsic plasticity, and structural plasticity. The memory flow: 1. Tokenize text into ID sequences; 2. Drive SORN to generate spike activity;3. Store patterns via plasticity rules;4. Retrieve via Jaccard similarity (token overlap);5. Inject relevant tokens into LLM prompts;6. Store LLM responses as new memory.

Key benefits: 1. Infinite memory capacity (compacted spike-weight patterns);2. Content-addressing (retrieve by topic similarity, not time);3. Constant inference cost (inject only 5-10 relevant tokens);4. No external embedding models (learns via STDP);5. Online continuous learning (updates weights immediately);6. Privacy (local storage) and explainability (clear token overlap triggers).

Use cases:1. Persistent chatbots (remember long-term conversations without increasing API costs);2. Biological memory research (model for studying STDP and synaptic plasticity);3. Privacy-sensitive fields (medical/finance with local storage);4. Low-resource environments (CPU-only operation).

Implemented in Julia (high-performance numerical computing). Core modules: SORNMemory.jl (main), bridge.jl (token-spike conversion), readout.jl (spike-token decoding), episodic_memory.jl (core API: store!/recall!/consolidate!), llm_interface.jl (NVIDIA NIM), context_injection.jl (prompt formatting), session.jl (chat loop), snn/ (SORN source). Example codes: verify_sorn.jl, pattern_learning.jl, chat_memory.jl (needs API key).

Current limitations: Only supports NVIDIA NIM API, in demo stage. Future plans: Support more LLM providers, optimize retrieval accuracy/recall, explore multi-modal memory, integrate with existing RAG frameworks.

SORNMemory represents a new paradigm for LLM memory, changing storage/retrieval via neuro-inspired SORN. Though experimental, it offers an exciting solution to long-term memory issues. It is based on academic research: Lazar et al. (2009) on SORN, Turrigiano & Nelson (2000) on homeostatic plasticity.