MedinaMemorySystems/HAM-384 (Hybrid Associative Memory)

HAM-384 is a highly compact embedding model designed for rapid memory retrieval, short-term to long-term memory (STM/LTM) consolidation, and enabling Hebbian spreading activation cascades in intelligent systems.

Model Details

Model Description

Fine-tuned from BAAI/bge-small-en-v1.5, HAM-384 is optimized for speed and associative recall. Its primary function is to act as the core vector retrieval mechanism in memory cascades, where memory nodes activate related nodes based on temporal and semantic proximity.

• Developed by: MedinaMemorySystems
• Model Type: Sentence Transformer / Feature Extraction
• Base Model: BAAI/bge-small-en-v1.5
• Language: English
• License: Apache-2.0

Vector Dimension & Retrieval Metrics

• Vector Dimension: 384
• MTEB Average: 51.8
• Pooling Strategy: Mean pooling

Architecture Details

• Architecture Type: BERT-small
• Parameters: 33 Million
• Layers: 6
• Attention Heads: 12
• Hidden Dimension: 384
• Max Sequence Length: 512 tokens

Intended Uses

• Primary Use Case: Vector representation for memory node retrieval in associative memory systems.
• Downstream Tasks:
  • Hebbian spreading activation cascades.
  • Hebbian synapse weight integration.
  • STM buffer encoding and temporal tick ordering.
  • Multi-hop resonance matching with a top-3 branching factor.

Quantization & Memory Footprint

Format	Precision	RAM Required	Latency (ms/query)
PyTorch	FP32 (Full)	~130 MB	~4.1
PyTorch	FP16	~65 MB	~2.5
ONNX	INT8	~33 MB	~1.3
GGUF	Q4_K_M	~20 MB	~0.8

Training Details

Training Data

The model was fine-tuned on associative memory traces:

• Sovereign LTM consolidation logs.
• Document-to-document co-activation pairs capturing contextual sequences.
• Temporal memory sequences reflecting realistic memory activation patterns.

Training Procedure

Optimized for associative recall and resonance matching rather than strict semantic equivalence, enabling the system to jump between temporally and contextually linked concepts.

Evaluation

Benchmark Results

HAM-384 sacrifices some general retrieval performance for highly specialized associative recall:

Benchmark	Metric	Score
MTEB Retrieval	Average	51.8
Memory-Recall@3 (Custom)	Accuracy	89.4%
Activation-Cascade (Custom)	F1-Score	82.1%
Peak Throughput	Docs/sec	~74,222

Note: Memory-Recall@3 and Activation-Cascade F1 are custom metrics specific to internal MedinaMemorySystems evaluations.

Usage

System Prompt / Prompting Template

For queries mimicking memory retrieval, no special prefix is enforced, though matching the style of associative logs helps recall:

{memory_fragment}

Example Code with SentenceTransformers

from sentence_transformers import SentenceTransformer, util

# Load the model
model = SentenceTransformer('MedinaMemorySystems/ham-384')

# A target memory to recall from
current_memory = "The user configured the security firewall rules for the cloud database."

# Memory bank (LTM)
memory_bank = [
    "Database backup completed successfully at midnight.",
    "Firewall updated to block unauthorized external IP ranges.",
    "User logged in from a new device."
]

# Encode the current state and memory bank
query_embedding = model.encode(current_memory)
corpus_embeddings = model.encode(memory_bank)

# Find top-3 nearest neighbors for spreading activation
hits = util.semantic_search(query_embedding, corpus_embeddings, top_k=3)[0]

print("Top Associative Retrievals:")
for hit in hits:
    print(f"- {memory_bank[hit['corpus_id']]} (Score: {hit['score']:.4f})")

Example Code with Transformers

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('MedinaMemorySystems/ham-384')
model = AutoModel.from_pretrained('MedinaMemorySystems/ham-384')

# Tokenize inputs
sentences = ["The user configured the security firewall rules."]
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform mean pooling
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0]
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)
print(sentence_embeddings)

Ethics & Safety

Ethics & Limitations

• General Retrieval: Due to its small size and specialization for associative cascades, it scores lower on general zero-shot retrieval tasks compared to larger models.
• Context Length: Truncates inputs longer than 512 tokens.
• Bias towards Co-occurrence: The model may retrieve documents based on temporal co-occurrence in its training data rather than pure semantic similarity.
• Safety: The model does not include specific safety filters or RLHF alignment; predictions are purely based on associative similarity from logs.

Citation

@misc{medinamemorysystems2026ham,
  title={HAM-384: A Compact BERT Model for Hybrid Associative Memory Cascades},
  author={MedinaMemorySystems},
  year={2026},
  publisher={Hugging Face}
}