Tommi Niemi
176164815b
Tommi Niemi / Rotko Networks Hidden-state episodic memory for frozen transformers. No gradients. Teach via one forward pass, recall via cosine similarity + logit injection. 200-line Python reproduction included. pip install transformers torch numpy && python python/epimem.py
97 lines
2.1 KiB
Plaintext
97 lines
2.1 KiB
Plaintext
// isis memory bank — FlatBuffers schema
|
||
// Supports f32/f16/i8 key quantization for production deployment
|
||
|
||
namespace isis.fb;
|
||
|
||
// Key quantization formats
|
||
enum KeyFormat : byte {
|
||
F32 = 0,
|
||
F16 = 1,
|
||
I8 = 2,
|
||
}
|
||
|
||
// Exact model identity — keys are ONLY valid for this exact config.
|
||
table ModelId {
|
||
model: string; // HuggingFace model name (e.g. "Qwen/Qwen2.5-0.5B")
|
||
backend: string; // Inference backend (e.g. "onnx", "gguf", "transformers")
|
||
quant: string; // Weight quantization (e.g. "f32", "f16", "q4_k_m")
|
||
hidden_dim: uint32; // Hidden state dimension (e.g. 896)
|
||
extraction: string; // Hidden state extraction point (e.g. "pre_mlp_layer23")
|
||
}
|
||
|
||
// A memory key stored in the chosen precision
|
||
table KeyData {
|
||
format: KeyFormat;
|
||
// Exactly one of these is populated based on format
|
||
f32_data: [float]; // dim × 4 bytes
|
||
f16_data: [uint16]; // dim × 2 bytes (IEEE 754 half)
|
||
i8_data: [int8]; // dim × 1 byte (scaled to [-127, 127])
|
||
// Scale factor for i8 dequantization: real = i8 * scale
|
||
i8_scale: float;
|
||
}
|
||
|
||
table SuppressEntry {
|
||
token_id: uint32;
|
||
bias: float;
|
||
}
|
||
|
||
table LogitBias {
|
||
token_id: uint32;
|
||
token: string;
|
||
strength: float;
|
||
suppress: [SuppressEntry];
|
||
}
|
||
|
||
table ContentKey {
|
||
key: KeyData;
|
||
token: string;
|
||
position: int32;
|
||
}
|
||
|
||
table Episode {
|
||
prompt: string;
|
||
answer: string;
|
||
alter: string;
|
||
keys: [ContentKey];
|
||
logit_biases: [LogitBias];
|
||
strength: float;
|
||
recall_count: uint32;
|
||
created_at: float64;
|
||
consolidated: bool;
|
||
}
|
||
|
||
table Alter {
|
||
name: string;
|
||
episodes: [Episode];
|
||
}
|
||
|
||
table Rule {
|
||
instruction: string;
|
||
priority: float;
|
||
trigger: string;
|
||
active: bool;
|
||
}
|
||
|
||
table Avoidance {
|
||
pattern: string;
|
||
reason: string;
|
||
key: KeyData;
|
||
suppress_token_ids: [uint32];
|
||
strength: float;
|
||
active: bool;
|
||
}
|
||
|
||
table MemoryBank {
|
||
version: uint32;
|
||
model_id: ModelId; // Exact model identity
|
||
threshold: float;
|
||
key_format: KeyFormat;
|
||
alters: [Alter];
|
||
rules: [Rule];
|
||
avoidances: [Avoidance];
|
||
}
|
||
|
||
root_type MemoryBank;
|
||
file_identifier "isis";
|
||
file_extension "fb";
|