Engineering Release Ledger

Changelog & Roadmap History

A chronological, verified record of all major architectural phases, performance baselines, and upcoming roadmap releases.

UnscheduledNo committed dateROADMAP TARGET

CoreVoiceInfrastructureUI

Roadmap: Not Yet Built

Items genuinely not built yet, with no target version or date committed -- see /roadmap for the full, itemized status of each (including why it isn't shipped).

Key Highlights

WebGPU in-browser inference: quantized small-LLM and ONNX voice synthesis entirely client-side. Exploratory -- nothing built or benchmarked yet.
1-click hosted Colab training runners for voice fine-tuning and benchmark sweeps, beyond the existing manual notebooks. Not started.
Community persona registry for publishing and browsing authored persona presets. Not started -- raises real moderation/consent/hosting questions not yet worked through.
Real GPT-SoVITS-rendered audio clips for the Voice Showcase, replacing the current static parameter table. In progress.

Target Client RAM (WebGPU)< 2.0 GB (unverified target)

Full itemized status on the Roadmap page →

COMING SOONNo committed date

Upcoming architectural milestone scheduled in the community roadmap.

v7.1.0September 2026

CoreUIMemoryVoice

Brain-First Repositioning: Live Playground, Verified Research, Repo Cleanup

The Brain (cognitive/affect/memory architecture) is now presented as the primary novelty, Voice and Vision as explicitly secondary. Every remaining fake 'Coming Soon' demo is replaced with real, live client-side math or moved to an honest /roadmap page, the homepage and showcase reorder around the Brain, and the Research page is rewritten around 5 pillars with every citation mechanically verified.

Key Highlights

Repo cleanup: outreach/, partnership/, orchestration/, and evidence/IP_REVIEW_CANDIDATES.md gitignored (internal BD/process material, not a history rewrite); FINAL_HUMANOID_BRAIN_ARCHITECTURE.md repurposed into a public ARCHITECTURE.md.
5 real backend formulas ported to TypeScript with unit tests against hand-computed values: ACT-R memory activation, Marsh trust + Bowlby attachment, metacognitive Brier-calibration, theory-of-mind concept tracking, and the persona compiler's _infer_temperament.
4 new live playground demos (Trust & Attachment, Memory Activation & Decay, Metacognitive Abstention, Theory of Mind) and a rewritten Persona Compiler that recomputes synchronously instead of faking a compile with setTimeout.
Homepage reordered to lead with the Brain (Endocrine -> Cognitive Turn -> Persona -> Trust -> Memory -> Mesh -> DevEx -> Voice -> Benchmarks -> Security -> CTA); the old fake roadmap-preview block is gone, replaced by a real /roadmap page.
Showcase recipes reordered to lead with Affective Dynamics, plus 2 new recipes (trust/Theory-of-Mind tracking, metacognitive honesty), Voice last.
Research page rewritten around 5 pillars with a new typed citation list; every citation mechanically checked against its claimed title/authors -- 6 of 35 had a real sourcing error (wrong author bylines, a wrong DOI digit, two stale arXiv IDs, a wrong ISBN) and were corrected, not dropped or left wrong.
Fixed: hero heading could overlap the fixed nav on a narrow-tall or short-landscape viewport (width-only breakpoints missed both cases) -- now a vmin-based clamp() with a short-viewport safety net.
Fixed: docs pages with LaTeX math rendered the raw $...$ text literally -- added remark-math + rehype-katex.
Fixed: this changelog's own unreleased-roadmap entry claimed a Home Assistant integration and peer-to-peer friend syncing with no grounding anywhere in the codebase -- replaced with the real, vetted /roadmap items.

Citations Verified35/35 (6 corrected)

Live Playground Demos8/8 (0 fake)

Vitest28/28 passing

v7.0.0August 2026

CoreUIVoiceMemoryBenchmarksInfrastructure

Community Release: Fresh-Clone Boot, Persona Compiler, Web UI & Distribution

The first tagged release of the community roadmap: a clean clone that boots and speaks, natural-language persona and voice creation, persistent affect and proactive outreach across restarts, four-store export/import, a real web UI, nine pressure-tested benchmark scenarios, and one-line installers with a unified friend CLI.

Key Highlights

Safety-floor fix: AppraisalEngine.identity_values now reads the same immutable_core['boundaries'] validate_response already used, closing a gap where it had silently been an empty list.
Fresh clone boots and speaks: bundled CC0 default voice sample, degrading GPT-SoVITS healthcheck, and a one-command start.sh across light/heavy/full/+vision profiles.
Natural-language persona compiler and 8-second voice enrollment with automated Whisper transcription and quality validation, plus a terminal chat REPL (scripts/talk.py).
Durable proactive outreach and affect state that survive a restart, with a reconnect queue for outreach missed while you were away.
Four-store export/import (Postgres, Neo4j, identity state, state_cache.db) and a provider-agnostic cloud LLM fallback (Anthropic/OpenAI/OpenRouter).
A real web UI: onboarding, WebSocket chat (/api/chat/ws), persona/voice/memory endpoints, all behind session auth.
The nine AUDIT.md pressure scenarios actually run against live infrastructure, not left as placeholders.
mypy, bandit, and the radon D/E/F complexity tier flipped from report-only to blocking in CI.
New this release: one-line installers for macOS/Linux and Windows, a ~4.3 MB standalone runtime bundle, the unified friend CLI (init/start/stop/status/model/vision/talk/persona/voice/backup/logs/update), an interactive friend init wizard generating cryptographically secure credentials, and Moondream VLM visual appraisal wired into setup.
Dedicated GPU Server & LiveKit WebRTC Voice Loop: real-time bidirectional voice over a physical SFU mesh with 39.95ms measured TTFT, zero-trust UFW firewall, and viseme data channel delivery.

Python Tests1,412 passing

Rust Tests130 passing

LLM TTFT (RTX 2060S)39.95 ms (Measured)

Peak RAM Headroom2.0 GB on 16GB Mac

v6.5.0May 2026

CoreMemoryBenchmarks

HippoRAG Retrieval & Mesh Architecture v3.5

A HippoRAG-style personalized PageRank retrieval engine over a co-occurrence graph, Qdrant and Neo4j interconnected for graph-based spreading activation, and the mesh architecture's v3.5 upgrade.

Key Highlights

HippoRAG PPR engine with co-occurrence graph edges and conditional seeding, plus ACT-R degree-scaled spreading-activation boosts.
Qdrant and Neo4j interconnected so spreading activation runs directly over the graph rather than vector similarity alone.
Cognitive appraisal (heuristic appraisal, APRA v2, SQLite ACT-R activation loops) migrated to the Rust cognitive-rust extension.
Mesh architecture v3.5: layered memory, attentional interruption, and loopback security hardening.

v6.0.0May 2026

CoreVoiceMemoryBenchmarks

Theory of Mind & Dynamic Prosody Mapping

Theory of Mind integration, continuous prosody trajectories with overlap-add crossfade, and neuromodulatory memory consolidation with dimensional trust.

Key Highlights

Theory of Mind integration (Phase 5): the agent models what the user does and doesn't know.
Dynamic Continuous Prosody Mapping with overlap-add (OLA) crossfade in the Rust voice agent.
Memory consolidation with neuromodulatory gating and dimensional trust scoring.
Physiology & sensory filters, plus a database-backed conversation store with automatic SQLite fallback.
Extended 8-dimensional cognitive benchmark suite with a compiled academic report.

v5.0.0May 2026

CoreVoiceMemoryInfrastructure

Rust Migration & Endocrine-Driven Sampling

Core cognitive agents migrated from Python to Rust, ACT-R memory storage and retrieval, and endocrine state (cortisol/dopamine) driving LLM sampling parameters for the first time.

Key Highlights

Python cognitive agents deprecated in favor of a hardened Rust mesh; compose profiles streamlined around it.
ActionService: endocrine-based LLM parameter control (cortisol/dopamine mapped to temperature/top_p) plus unified NATS SubjectMetrics.
ACT-R memory storage and retrieval with integration testing.
SenseVoice STT with emotional perception and paralinguistics tracking.
Global Python stack standardized to 3.12 across Dockerfiles and CI.

v4.1.5May 2026

InfrastructureSecurity

Release Pipeline Hardening

Conventional Commits and SBOM dependency graphs, signed release assets, and cross-platform (Windows/macOS) build matrix fixes.

Key Highlights

Enterprise-grade release pipeline: Conventional Commits, SBOM dependency graphs, and versioned manifests.
Release assets now include SHA256 checksums, an offline CHANGELOG, LICENSE, and metadata.json.
Fixed a macOS package suffix mismatch and upgraded CycloneDX to a compliant v1.4 output.
Fixed a Windows Compress-Archive positional-argument bug by separating matrix shell targets.

v4.0.0May 2026

CoreVoiceSecurity

Endocrine System & Subconscious Processing

The endocrine simulation (cortisol/dopamine modulation of LLM generation parameters) and agent-based subconscious background processing land for the first time.

Key Highlights

Endocrine system: cortisol and dopamine modulation of LLM generation parameters.
Agent-based subconscious processing modules for background reflection.
GPT-SoVITS orchestration, service configuration, and identity warmup scenarios updated.
CI/CD suite expanded with security auditing and infrastructure validation workflows.

v3.2.4April 2026

CoreVoiceMemoryInfrastructure

AI Friend Core: Solid State Mesh

Zero-drift NATS micro-agent architecture: BaseAgent, BrainAgent, VoiceAgent, and Neo4j-backed cognitive memory, plus GPT-SoVITS voice cloning setup.

Key Highlights

AI Friend Core Solid State Mesh: zero-drift NATS-based micro-agent architecture with a shared BaseAgent.
BrainAgent with semantic chunking; VoiceAgent with adaptive audio normalization and state-managed synthesis.
Neo4j-backed cognitive memory and a SurfacingAgent for asynchronous long-term memory retrieval.
GPT-SoVITS voice cloning setup, training documentation, and a voice-training notebook.

v3.0.0February 2026

CoreVoiceInfrastructure

Agent-Based Cognitive Mesh

The first cognitive mesh architecture: independent micro-agents, graph RAG, and a NATS event bus, with WebRTC voice and vision capabilities.

Key Highlights

Agent-based architecture with WebRTC voice and vision capabilities.
v3.0 cognitive mesh: micro-agents, graph RAG, and a NATS event bus.
CI workflow for linting, security analysis, frontend build, and backend import validation.

v2.2.0February 2026

CoreSecurityInfrastructure

The Enterprise Era

CI/CD and security hardening on top of the imperfect-memory and dynamic-identity core: automated Docker publishing, dependency security enforcement, and broken-link checking.

Key Highlights

Automated Docker image publishing to GHCR alongside CI testing/validation and tag-triggered releases.
Long-term memory, vision, and audio processing folded into a new interactive assistant UI.
Dependency security enforcement pass across requirements.txt.
Broken-link checking workflow for markdown documentation.

v2.0.0January 2026

CoreMemory

The Humanized Era

The project's rebrand to AI Friend, and the first version of imperfect memory, a simulated biological clock, and identity driven by configuration rather than hardcoded prompts.

Key Highlights

Imperfect Memory (Memory v2): short-term, blurry, and core memory layers.
Biological Clock: temporal awareness and simulated biological rhythms.
Dynamic Identity: personality and history derived from JSON configuration rather than hardcoded prompts.
Human Growth Engine: session reflection and persistent learning.
Rebranded the project to AI Friend, with a dynamic LLM service and centralized configuration.