SORNMemory — SORN-powered episodic memory for LLMs. Demo project.

A self-organizing spiking neural network (SORN) that stores chat history as temporal spike patterns and retrieves relevant context for LLM prompts. The idea of using SORN as a temporal, self-organizing memory module for LLMs is unique to this research direction.

Setup

git clone https://github.com/omgbox/SORNMemory.git
cd SORNMemory
julia --project=. -e 'using Pkg; Pkg.instantiate()'

API Key

Create keys.txt in the project root:

nvida Nim key: nvapi-your-key-here

Only NVIDIA NIM is supported.

Run

No API key needed:

julia --project=. examples/verify_sorn.jl
julia --project=. examples/pattern_learning.jl

With API key:

julia --project=. examples/chat_memory.jl

Commands inside chat: quit, stats, history, quiet, verbose

Benefits

• Beyond context windows — LLM context limits are fixed (8K-128K tokens). SORN stores unlimited conversation history as compressed spike-weight patterns, not raw text. No sliding window, no truncation.
• Content-addressable recall — Retrieves past episodes by topical overlap (Jaccard similarity on token IDs), not recency. A question about something said 500 messages ago triggers the same recall as something said 5 messages ago.
• Inference-time cost efficiency — Instead of prepending the entire conversation history (growing token costs linearly with each turn), SORN injects only ~5-10 relevant token IDs. Cheaper and faster per call.
• No pretrained embeddings — No dependency on BERT, SentenceTransformers, or any external embedding model. The SORN network learns temporal patterns purely through spike-timing-dependent plasticity.
• Online learning — Every message updates the network immediately. No batch training, no finetuning, no gradient descent.

Use Cases

• Persistent chatbots — Give an LLM long-term memory beyond its context window without ballooning API costs.
• Research on biological memory — Study how STDP, synaptic scaling, and intrinsic plasticity interact to store and recall sequences.
• Privacy-sensitive applications — All memory is in local SORN weight matrices, not in a cloud vector database.
• Low-resource memory — SORN runs on CPU (375 neurons, 13K synapses). No GPU needed.
• Explainable memory retrieval — Every recall shows exactly which token overlap triggered the match, unlike opaque vector embedding similarity.

How It Works

1. Each message is tokenized by a bidirectional word-level tokenizer (word → ID, ID → word). Unknown words are assigned new IDs on the fly up to a configurable vocabulary size.
2. Token IDs drive a 375-neuron SORN with 5 plasticity rules (STDP, ISTDP, synaptic scaling, intrinsic plasticity, structural plasticity)
3. On recall, the current message tokens are matched against stored episodes by Jaccard similarity
4. The best-matching episode tokens are decoded back to words and injected as context into the LLM prompt
5. The LLM response is also tokenized and stored back into SORN as a new episode

Example chat output showing the tokenizer in action:

You: What about Zephyrian quantum entanglement in the Naxos protocol?
[SORN] Tokens: [262, 227, 344, 345, 346, 18, 213, 347, 348]
            -> [what, about, zephyrian, quantum, entanglement, in, the, naxos, protocol]
[SORN] Recalled: 344 (zephyrian), 345 (quantum), 346 (entanglement)
[SORN] Injected: "zephyrian", "quantum", "entanglement"

Requirements

• Julia 1.10+
• HTTP.jl, JSON3.jl
• NVIDIA NIM API key

Project Structure

SORNMemory/
├── src/
│   ├── SORNMemory.jl          # Main module
│   ├── tokenizer.jl           # Bidirectional word-level tokenizer (online vocab growth)
│   ├── bridge.jl              # Token ↔ spike encoding
│   ├── readout.jl             # Spike pattern → token decoding
│   ├── episodic_memory.jl     # store!/recall!/consolidate! — core memory API
│   ├── llm_interface.jl       # NVIDIA NIM provider
│   ├── context_injection.jl   # Memory → prompt formatting
│   ├── session.jl             # Chat loop orchestrator
│   └── snn/                   # Bundled SORN neural network source
├── examples/
│   ├── verify_sorn.jl         # SORN pipeline verification (no API key)
│   ├── pattern_learning.jl    # SORN learning demo (no API key)
│   └── chat_memory.jl         # Full chat with SORN memory
├── keys.template.txt          # API key format template
├── LICENSE                    # MIT License
├── Project.toml
└── README.md

References

• Lazar et al. (2009) "SORN: a self-organizing recurrent neural network"
• Turrigiano & Nelson (2000) "Homeostatic plasticity in developing networks"