================================================================================
CHANGELOG — titans-torch
================================================================================

All notable changes to this project are documented in this file.
Format follows: [Version] YYYY-MM-DD — summary of changes.

--------------------------------------------------------------------------------
[0.1.0] 2025-04-03  —  Initial Release
--------------------------------------------------------------------------------

Added
  - NeuralLongTermMemory (LMM)
      · forward_trainable(): differentiable inner-loop chunkwise gradient
        descent, suitable for end-to-end outer-loop training.
      · forward(): pure inference pass with no memory update.
      · test_time_update(): sequential per-token adaptation at test time
        without retaining a computation graph.
      · Data-dependent learning-rate coefficients (alpha, theta, eta) via
        small MLP nets.
      · Optional 1D depthwise conv on K/V/Q projections.
      · Momentum-based inner optimizer: S_t = eta*S_{t-1} - theta*grad.

  - PersistentMemory
      · Learnable token bank prepended to any input sequence.
      · Configurable number of tokens and initialisation scale.

  - TitansMAC (Memory as a Context)
      · Segment-wise retrieval from long-term memory.
      · Attention over [persistent ‖ retrieved ‖ segment].
      · Post-attention memory update followed by gated output.
      · Configurable num_heads and dropout.

  - TitansMAG (Memory as a Gate)
      · Parallel sliding-window short-term attention and LMM.
      · Learned gate combines both streams.
      · Configurable window_size, num_heads, dropout.

  - TitansMAL (Memory as a Layer)
      · LMM applied first as a preprocessing layer.
      · Sliding-window attention applied on top.
      · Configurable window_size, num_heads, dropout.

  - Package scaffolding
      · setup.py, MANIFEST.in, README.txt, LICENSE.txt, CHANGELOG.txt.

--------------------------------------------------------------------------------
[Unreleased]
--------------------------------------------------------------------------------

Planned / under consideration
  - Parallel associative scan for momentum recurrence (speedup).
  - Multi-layer stacking utility (e.g. TitansStack).
  - FP16 / BF16 mixed-precision support and benchmarks.
  - Integration tests and example training scripts.
  - PyPI release.