epimem: One-shot gradient-free learning on frozen transformers

- paper.md: full paper (Tommi Niemi / Rotko Networks)
- python/epimem.py: standalone Python reproduction
- export_onnx.py: ONNX export from HuggingFace (generates model files)
- results/memory_bank.json: example hidden-state vectors (896-dim)
- schema/: FlatBuffer schemas for memory bank + organism
- models/tokenizer/: Qwen 2.5 tokenizer files

Run: pip install transformers torch && python python/epimem.py
(Downloads Qwen 2.5 automatically from HuggingFace)

schema/isis.fbs

@@ -0,0 +1,96 @@
+// 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";

schema/organism.fbs

@@ -0,0 +1,257 @@
+// Organism schema — FlatBuffers
+//
+// Separates:
+//   Weights  (shared cortical hardware, Arc-shareable)
+//   Session  (per-persona state, isolated)
+//   Memory   (episodic store with per-alter access control)
+//
+// File layout:
+//   organism.weights.fb  — one per organism (shared across personas)
+//   persona_X.session.fb — one per persona
+//   memory.fb            — shared episodic store (existing isis.fbs)
+//   barriers.fb          — dissociative access control
+
+namespace isis.organism;
+
+// ─── Shared types ────────────────────────────────────────
+
+enum Precision : byte {
+    F32 = 0,
+    F16 = 1,
+    I8  = 2,
+}
+
+/// Dense matrix stored in chosen precision.
+/// All weight matrices use this: synapses, NLM stages, projectors.
+table Matrix {
+    rows: uint32;
+    cols: uint32;
+    precision: Precision;
+    f32_data: [float];
+    f16_data: [uint16];
+    i8_data: [int8];
+    i8_scale: float;
+}
+
+/// Bias vector.
+table Bias {
+    precision: Precision;
+    f32_data: [float];
+    f16_data: [uint16];
+    i8_data: [int8];
+    i8_scale: float;
+}
+
+// ─── Region definition ───────────────────────────────────
+
+/// Per-neuron MLP weights within a region.
+table NlmStage {
+    weights: Matrix;     // [n_neurons × out_per × in_per]
+    biases: Bias;        // [n_neurons × out_per]
+    n_neurons: uint32;
+    in_per: uint32;
+    out_per: uint32;
+}
+
+/// A brain region's learned weights.
+table RegionWeights {
+    name: string;        // "input", "attention", "output", "motor", etc.
+    n_neurons: uint32;
+    memory_length: uint32;
+    inhibitory_fraction: float;
+    inhibitory_mask: [bool];
+    nlm_stage1: NlmStage;
+    nlm_stage2: NlmStage;   // null if nlm_depth < 2
+    start_trace: [float];   // initial trace state
+    start_activated: [float]; // initial activation
+}
+
+/// Inter-region synapse weights.
+table SynapseWeights {
+    from_region: string;
+    to_region: string;
+    weight: Matrix;      // [out_dim*2 × in_dim] (×2 for GLU)
+    bias: Bias;          // [out_dim*2]
+}
+
+// ─── Weights file (shared across personas) ───────────────
+
+table OrganismWeights {
+    version: uint32;
+
+    // Architecture config
+    iterations: uint32;
+    d_model: uint32;
+    d_input: uint32;
+    n_sync_out: uint32;
+    n_sync_action: uint32;
+    motor_threshold: float;
+
+    // Brain regions (variable count — not hardcoded to 8)
+    regions: [RegionWeights];
+
+    // Inter-region synapses (variable count)
+    synapses: [SynapseWeights];
+
+    // Projectors
+    global_projector: Matrix;
+    global_projector_bias: Bias;
+    output_projector: Matrix;
+    output_projector_bias: Bias;
+    logit_projector: Matrix;
+    logit_projector_bias: Bias;
+
+    // Sync pair topology (deterministic from seed, but stored for portability)
+    sync_out_left: [uint32];
+    sync_out_right: [uint32];
+    sync_out_decay: [float];
+    sync_action_left: [uint32];
+    sync_action_right: [uint32];
+    sync_action_decay: [float];
+
+    // Position predictor weights (optional)
+    position_predictor_gate: Matrix;
+    position_predictor_content: Matrix;
+    position_predictor_final: Matrix;
+
+    // Organism-level learned parameters
+    embeddings: Matrix;         // [vocab_size × embed_dim]
+    sensory_layers: [Matrix];   // sensory MLP weight matrices
+    sensory_biases: [Bias];
+    output_proj: Matrix;        // [vocab_size × n_sync_out]
+    output_proj_bias: Bias;
+
+    // Metadata
+    vocab_size: uint32;
+    embed_dim: uint32;
+    sensory_depth: uint32;
+    context_len: uint32;
+    tokens_seen: uint64;
+    sleep_cycles: uint64;
+    created_at: float64;
+}
+
+// ─── Session file (per-persona) ──────────────────────────
+
+/// Neuromodulator baseline — each persona's emotional temperament.
+table NeuromodBaseline {
+    dopamine: float;
+    serotonin: float;
+    norepinephrine: float;
+    acetylcholine: float;
+    curiosity: float;
+    anxiety: float;
+}
+
+/// Hippocampal content-addressable memory state.
+table HippocampalState {
+    capacity: uint32;
+    key_dim: uint32;
+    value_dim: uint32;
+    keys: [float];        // [capacity × key_dim]
+    values: [float];      // [capacity × value_dim]
+    strengths: [float];   // [capacity]
+    write_ptr: uint32;
+    count: uint32;
+}
+
+/// Per-region mutable state (noise, usefulness tracking).
+table RegionState {
+    region_name: string;
+    noise_scale: [float];      // [n_neurons]
+    usefulness_ema: [float];   // [n_neurons]
+}
+
+/// Hebbian plasticity state per region.
+table HebbianState {
+    region_name: string;
+    running_mean: [float];     // [n_neurons]
+    baseline_mean: [float];
+    baseline_var: [float];
+    calibrated: bool;
+}
+
+/// Replay buffer entry for prioritized consolidation.
+table ReplayEntry {
+    observation: [float];
+    surprise: float;
+    timestamp: uint64;
+}
+
+/// A persona's complete session state.
+/// This is everything that makes one persona different from another
+/// running on the same shared weights.
+table PersonaSession {
+    version: uint32;
+
+    // Identity
+    name: string;              // persona name
+    created_at: float64;
+
+    // Emotional temperament
+    neuromod_baseline: NeuromodBaseline;
+
+    // Episodic memory (per-persona hippocampal state)
+    hippo: HippocampalState;
+    hippo_retrieval: [float];
+
+    // Basal ganglia learning state
+    bg_eligibility: [float];
+    bg_da_baseline: float;
+    bg_weight_delta: [float];
+
+    // Per-region mutable state
+    region_states: [RegionState];
+
+    // Per-region Hebbian state
+    hebbian_states: [HebbianState];
+
+    // Sleep consolidation traces (transient, may be empty)
+    sleep_trace_count: uint32;
+
+    // Replay buffer
+    replay_entries: [ReplayEntry];
+    replay_capacity: uint32;
+
+    // Training state
+    hebbian_enabled: bool;
+    tokens_processed: uint64;
+    gamma_cycles_total: uint64;
+}
+
+// ─── Dissociative barriers ───────────────────────────────
+
+/// Per-persona access control over shared episodic memory.
+/// Each persona can see different episodes with different strength.
+table EpisodeAccess {
+    episode_index: uint32;     // index into MemoryBank.alters[].episodes[]
+    access_level: float;       // 0.0 = walled off, 1.0 = full access
+}
+
+table PersonaBarrier {
+    persona_name: string;
+    episode_access: [EpisodeAccess];
+}
+
+table DissociativeBarriers {
+    version: uint32;
+    barriers: [PersonaBarrier];
+    // Global integration level: 0.0 = fully dissociated, 1.0 = fully integrated
+    integration_level: float;
+}
+
+// ─── Root types ──────────────────────────────────────────
+
+root_type OrganismWeights;
+file_identifier "orgw";
+file_extension "weights.fb";
+
+// For session files, use PersonaSession as root:
+// root_type PersonaSession;
+// file_identifier "orgs";
+// file_extension "session.fb";
+
+// For barrier files, use DissociativeBarriers as root:
+// root_type DissociativeBarriers;
+// file_identifier "orgb";
+// file_extension "barriers.fb";