Changelog

[0.5.0] - 2026-09-27

Stage 11: probing memory

  • Probe (side panel, or Shift-click a row): a trace’s address used as a cue, for inspection. Every trace is ranked by its activation, with its similarity under the current Address and under each address set; the matrix shows the probe’s activations.
  • Play echo plays the probe’s echo soloed, once or looped: the raw blend of the answering traces, or through the echo’s tone, tape, level and spring (Full path). Click a ranked trace to hear it on its own.
  • Compare with traces’ own addresses (n vs n) or with their context (n vs n−1: what memory expects next); Include itself.
  • Everything updates as retrieval settings change, even while the echo loops.
  • Probing never changes memory, the random streams or the plug-in’s output (a test renders with and without probing and compares).

[0.4.0] - 2026-09-27

Stage 10: address sets

  • Every trace stores five addresses: Spectrum (the original), Pitch Class (chroma, octave-blind), Pitch (semitones C2–B5), Timbre (spectral shape and brightness, aligned to the fundamental) and Rhythm (onsets in four frequency ranges). The Address parameter chooses which one memory compares; Custom mixes them with five weights (weighted mean of the sets’ similarities). Sequence context stores every set too.
  • With a pitch address the echo chain follows melodies: a scale learned one note per bar is walked back in order (with the Spectrum, neighbouring notes are 0.98 similar and the walk drifts).
  • The memory matrix and side panel show the set being compared, in its own layout; new Address parameter page.
  • Memory files store every set; memory saved by earlier versions gets the new sets computed from its audio. Older versions still read new files.
  • The default (Spectrum) renders exactly as before: every factory preset and earlier listening example is bit-identical.
  • Echo Chain switched on after the input has stopped now starts from the echo that is playing (it used to wait for input).
  • Factory presets Scale Walker, Melody Memory, Same Sound, Groove Recall; listening examples 38–42 and a scale test signal; an Addresses page in the manual.

[0.3.0] - 2026-09-27

  • Running / Paused (header button and parameter): pausing leaves memory untouched (nothing stored, forgotten or cued), fades the echo out and passes the dry signal. Resuming starts cleanly at the next segment.
  • Audition: listen to a trace on its own (side panel, or Alt-click a row), soloed over the plug-in’s output, running or paused; Loop; and n−1 | n to hear the segment before it first. Never recorded.
  • Chain Step: Sample (new default) steps the echo chain to one answering trace drawn by activation, a random walk through memory that follows learned transitions; Blend keeps the previous deterministic behaviour. Fixes Playback = Sample not affecting the walk.
  • Cue Noise: noise on every cue (live, chain, iterative, progressive).
  • Habituation: traces that just answered are briefly less active.
  • Dreaming Sequencer and example 36 use the sampled walk.

[0.2.0] - 2026-09-27

Stage 9: sequential context

After Jamieson and Mewhort (2009), who extended MINERVA to the serial reaction-time task, and in the spirit of Elman’s (1990) context units.

  • Every trace also stores the previous segment’s address: [n−1 | n]. Trace audio is unchanged. Context is recorded always, saved with memory (an optional context field; older files load without it) and built for imported audio from the file’s previous piece.
  • Sequence Context and Context Cue: Predict Next (the segment just played against traces’ n−1 halves: memory anticipates the next segment), Match Both (sequence-sensitive recall, with Context Weight) or Current Only. Works with Segment and Progressive cueing and iterative heads.
  • Echo Chain cue source: each echo’s expected next segment cues the next echo, so memory walks through learned sequences. Chain Input blends in the live input (0 = free-running). MIDI base note + 8 holds the chain.
  • Memory matrix and side panel show [n−1 | n]; new Sequence page.
  • Factory presets Anticipate, Sequence Memory, Dreaming Sequencer; examples 34–37; a Sequences page in the manual.
  • Off by default: with Sequence Context off, output is unchanged.

[0.1.0] - 2026-09-27

First release: every stage of plan.md is in place. It has been tested offline and in automated validators (auval, pluginval), but not yet played in Live, so expect rough edges.

Stage 8: polish and release

  • Blocks larger than the size the host announced (offline bounces) are split inside the engine instead of overrunning its buffers.
  • Changing the sample rate or channel count keeps memory (resampled or mixed down) instead of clearing it.
  • Non-finite input (NaN/Inf from a misbehaving plug-in upstream) is replaced by silence before it can reach memory.
  • Tempo changes mid-segment, transport loops, odd and zero-length blocks and determinism of offline renders are covered by tests.
  • Parameter smoothing audit: every continuous parameter is switched back and forth during a steady tone and checked for clicks (now a test). Wow, flutter, spring level, echo tone, tape drive, feedback, the feedback shelves and edge fade used to click; they now glide.
  • Optimization: vectorised FFT (2.7x faster), phases taken once per bin in the spectral engine, a fast path for plain playback and vectorised similarity. Most presets run 1.5-2x faster; the spectral ones about 2x (26-30x realtime, from 14x).
  • macOS installer (.pkg) alongside the zip; optional Developer ID signing and notarization in CI; tagged versions become GitHub releases.

Stage 7: custom UI and control

  • Memory matrix editor (traces as rows with their addresses, activations, heads and locks), Heard/Echo spectrograms, familiarity meter, trace inspector, tabbed parameter pages.
  • Drag-and-drop audio import, MIDI note control, trigger parameters, and factory, example and user presets.

Stage 6: spectral engine

  • Spectral blending, time-stretch for length mismatch, spectral freeze, grain memories of up to 4000 traces.

Stage 5: tape character

  • Varispeed, wow and flutter, tape drive, hiss, feedback bass/treble and a spring reverb.

Stage 4: heads and chorus

  • RE-201-style playback heads and mode selector, chorus voices, sampled playback, sidechain, random and frozen cues, recency and feature focus.

Stage 3: live cueing

  • Progressive (within the bar) and rolling (no bar grid) cueing.

Stage 2: memory management

  • Freezing, write gates, clamping, full-memory policies, consolidation, forgetting, echo re-encoding, and saving and loading memory.

Stage 1: tape = memory

  • Traces stored per segment, cued by the segment just played; the echo is the activation-weighted blend. Memory Capacity = 1 is a tape delay.

Stage 0: scaffolding

  • JUCE/CMake project, engine library, tests, offline tools, CI.