What's New in v2.1.23 "Receipt Lock Hardening"

v2.1.23 turns the Sanhedrin Receipt Lock launch into something more portable, observable, and harder to spoof.

• Model-agnostic Sanhedrin presets. Sanhedrin no longer guesses a large default verifier. Users choose any OpenAI-compatible endpoint/model, or start from custom, small laptop, Ollama, MLX, vLLM, llama.cpp, hosted API, or LiteLLM presets.
• Sharper Receipt Lock. Verification claims inside code fences, quotes, blockquotes, or explicitly hedged "let me verify" language no longer trigger false vetoes, while actual "tests passed" claims still require command receipts.
• Safer command receipts. Transcript command evidence now prefers structured tool-use receipts; loose JSON scanning is opt-in only.
• Visible fail-open telemetry. Timeouts, unavailable model endpoints, and malformed verdicts are logged locally and surfaced in the dashboard's 7-day Sanhedrin stats.
• Durable evidence boundary. Staged evidence remains useful context, but it cannot satisfy durable support or contradiction requirements by itself.
• Safer batch writes. smart_ingest batch mode now keeps caller-separated items separate by default and returns merge previews when an existing memory is mutated.
• Opt-in NVIDIA acceleration path. Qwen3 embedding builds expose CUDA/cuDNN feature flags for contributors and users with CUDA-capable hosts.

Quick Start

# 1. Install
npm install -g vestige-mcp-server@latest

# 2. Connect to any MCP-compatible agent
# Claude Code
claude mcp add vestige vestige-mcp -s user

# Codex
codex mcp add vestige -- vestige-mcp

# OpenCode
npx @vestige/init

# 3. Test it
# "Remember that I prefer TypeScript over JavaScript"
# ...new session...
# "What are my coding preferences?"
# → "You prefer TypeScript over JavaScript."

vestige update

vestige update --install-dir /usr/local/bin

brew install onnxruntime
npm install -g vestige-mcp-server@latest
echo 'export ORT_DYLIB_PATH="'"$(brew --prefix onnxruntime)"'/lib/libonnxruntime.dylib"' >> ~/.zshrc
source ~/.zshrc
claude mcp add vestige vestige-mcp -s user

npm install -g vestige-mcp-server@latest
vestige-mcp --version

{
  "mcpServers": {
    "vestige": {
      "command": "vestige-mcp"
    }
  }
}

git clone https://github.com/samvallad33/vestige && cd vestige
cargo build --release -p vestige-mcp

npm install -g vestige-mcp-server

git clone https://github.com/samvallad33/vestige && cd vestige
cargo build --release -p vestige-mcp
# Optional: enable Metal GPU acceleration on Apple Silicon
cargo build --release -p vestige-mcp --features metal

Works Everywhere

Vestige speaks MCP, so any client that can register a stdio MCP server can use it.

🧠 3D Memory Dashboard

Vestige v2.0 ships with a real-time 3D visualization of your AI's memory. Every memory is a glowing node in 3D space. Watch connections form, memories pulse when accessed, and the entire graph come alive during dream consolidation.

Features:

• Force-directed 3D graph with 1000+ nodes at 60fps
• Bloom post-processing for cinematic neural network aesthetic
• Real-time WebSocket events: memories pulse on access, burst on creation, fade on decay
• Dream visualization: graph enters purple dream mode, replayed memories light up sequentially
• FSRS retention curves: see predicted memory decay at 1d, 7d, 30d
• Command palette (Cmd+K), keyboard shortcuts, responsive mobile layout
• Installable as PWA for quick access

Tech: SvelteKit 2 + Svelte 5 + Three.js + Tailwind CSS 4 + WebSocket

Run vestige dashboard to open http://localhost:3927/dashboard, or set VESTIGE_DASHBOARD_ENABLED=true to start it with the MCP server.

Architecture

┌─────────────────────────────────────────────────────┐
│  SvelteKit Dashboard (apps/dashboard)                │
│  Three.js 3D Graph · WebGL + Bloom · Real-time WS   │
├─────────────────────────────────────────────────────┤
│  Axum HTTP + WebSocket Server (port 3927)            │
│  15 REST endpoints · WS event broadcast              │
├─────────────────────────────────────────────────────┤
│  MCP Server (stdio JSON-RPC)                         │
│  25 tools · 30 cognitive modules                     │
├─────────────────────────────────────────────────────┤
│  Cognitive Engine                                    │
│  ┌──────────┐ ┌──────────┐ ┌───────────────┐       │
│  │ FSRS-6   │ │ Spreading│ │ Prediction    │       │
│  │ Scheduler│ │ Activation│ │ Error Gating  │       │
│  └──────────┘ └──────────┘ └───────────────┘       │
│  ┌──────────┐ ┌──────────┐ ┌───────────────┐       │
│  │ Memory   │ │ Synaptic │ │ Hippocampal   │       │
│  │ Dreamer  │ │ Tagging  │ │ Index         │       │
│  └──────────┘ └──────────┘ └───────────────┘       │
├─────────────────────────────────────────────────────┤
│  Storage Layer                                       │
│  SQLite + FTS5 · USearch HNSW · Nomic Embed v1.5    │
│  Optional: Nomic v2 MoE · Qwen3 Reranker · Metal   │
└─────────────────────────────────────────────────────┘

Why Not Just Use RAG?

RAG is a dumb bucket. Vestige is an active organ.

🔬 The Cognitive Science Stack

This isn't a key-value store with an embedding model bolted on. Vestige implements real neuroscience:

Prediction Error Gating — The hippocampal bouncer. When new information arrives, Vestige compares it against existing memories. Redundant? Merged. Contradictory? Superseded. Novel? Stored with high synaptic tag priority.

FSRS-6 Spaced Repetition — 21 parameters governing the mathematics of forgetting. Frequently-used memories stay strong. Unused memories naturally decay. Your context window stays clean.

HyDE Query Expansion (v2.0) — Template-based Hypothetical Document Embeddings. Expands queries into 3-5 semantic variants, embeds all variants, and searches with the centroid embedding for dramatically better recall on conceptual queries.

Synaptic Tagging — A memory that seemed trivial this morning can be retroactively tagged as critical tonight. Based on Frey & Morris, 1997.

Spreading Activation — Search for "auth bug" and find the related JWT library update from last week. Memories form a graph, not a flat list. Based on Collins & Loftus, 1975.

Dual-Strength Model — Every memory has storage strength (encoding quality) and retrieval strength (accessibility). A deeply stored memory can be temporarily hard to retrieve — just like real forgetting. Based on Bjork & Bjork, 1992.

Memory Dreaming — Like sleep consolidation. Replays recent memories to discover hidden connections, strengthen important patterns, and synthesize insights. Dream-discovered connections persist to a graph database. Based on the Active Dreaming Memory framework.

Waking SWR Tagging — Promoted memories get sharp-wave ripple tags for preferential replay during dream consolidation. 70/30 tagged-to-random ratio. Based on Buzsaki, 2015.

Autonomic Regulation — Self-regulating memory health. Auto-promotes frequently accessed memories. Auto-GCs low-retention memories. Consolidation triggers on 6h staleness or 2h active use.

Active Forgetting (v2.0.5) — Top-down inhibitory control via the suppress tool. Other memory systems implement passive decay — the Ebbinghaus 1885 "use it or lose it" curve, sometimes with trust-weighted strength factors. Vestige v2.0.5 also implements active top-down suppression: each suppress call compounds (Suppression-Induced Forgetting, Anderson 2025), a background Rac1 cascade worker fades co-activated neighbors across the connection graph (Cervantes-Sandoval & Davis 2020), and a 24-hour labile window allows reversal (Nader reconsolidation semantics on a pragmatic axis). The memory persists — it's inhibited, not erased. Explicitly distinct from Anderson 1994 retrieval-induced forgetting (bottom-up, passive competition during retrieval), which is a separate, older primitive that several other memory systems implement. Based on Anderson et al., 2025 and Cervantes-Sandoval et al., 2020. First shipped AI memory system with this primitive.

Full science documentation ->

🛠 MCP Tools

Context Packets

Core Memory

Cognitive Engine

Autonomic

Scoring & Dedup

Maintenance

Deep Reference (v2.0.4)

Active Forgetting (v2.0.5)

Make Your AI Use Vestige Automatically

Registering the MCP server exposes tools; the agent still needs an instruction that tells it when to call memory. Use the agent-neutral protocol, then adapt it to your client-specific instruction file.

Agent memory protocol -> · Claude Code template ->

Technical Details

Optional Features

# Qwen3 embeddings (Candle backend; add metal on Apple Silicon)
cargo build --release -p vestige-mcp --features qwen3-embeddings,metal
VESTIGE_EMBEDDING_MODEL=qwen3-0.6b vestige consolidate

Building with CUDA support (NVIDIA hosts - Windows / Linux)

The cuda feature routes Qwen3 embedding through NVIDIA GPUs via candle-core/cuda. On a host with the CUDA toolkit installed and a supported NVIDIA runtime, this drops Qwen3-Embedding inference from CPU-bound to GPU-bound for batched workloads.

# Linux / Windows + CUDA toolkit (12.x or 13.x)
cargo build --release -p vestige-mcp --features qwen3-embeddings,cuda

# Optional cuDNN acceleration on top of CUDA
cargo build --release -p vestige-mcp --features qwen3-embeddings,cudnn

VESTIGE_EMBEDDING_MODEL=qwen3-0.6b vestige consolidate

Prerequisites:

• NVIDIA driver + CUDA toolkit (12.x or 13.x). Verify with nvcc --version.
• A C++ host compiler that nvcc can drive (Linux: gcc; Windows: MSVC / cl.exe from a recent Visual Studio Build Tools install).

Windows + MSVC + CUDA 13.x build note. Recent CCCL headers shipped with CUDA 13.x require the modern preprocessor. Without it, the candle-kernels .cu compile pass can fail at cuda/include/cuda/std/__cccl/compiler.h. Set this env var before cargo build to pass /Zc:preprocessor through nvcc:

# PowerShell
$env:NVCC_PREPEND_FLAGS = '-Xcompiler="/Zc:preprocessor"'
cargo build --release -p vestige-mcp --features qwen3-embeddings,cuda

:: cmd.exe
set NVCC_PREPEND_FLAGS=-Xcompiler="/Zc:preprocessor"
cargo build --release -p vestige-mcp --features qwen3-embeddings,cuda

Linux + CUDA 13.x builds with gcc do not need the equivalent flag.

Verifying GPU is actually used. With CUDA-enabled builds, run VESTIGE_EMBEDDING_MODEL=qwen3-0.6b vestige consolidate on a corpus of 1000+ memories and watch nvidia-smi; embedding passes should pin a single GPU while the run is active.

CLI

vestige stats                    # Memory statistics
vestige stats --tagging          # Retention distribution
vestige stats --states           # Cognitive state breakdown
vestige health                   # System health check
vestige consolidate              # Run memory maintenance
vestige restore <file>           # Restore from backup
vestige portable-export <file>         # Exact cross-device archive
vestige portable-import <file>         # Import archive into an empty database
vestige portable-import <file> --merge # Merge archive into this database
vestige sync <file>                    # Pull/merge/push via file backend
vestige dashboard                # Open 3D dashboard in browser

Documentation

Troubleshooting

which vestige-mcp
# Or use the full path:
claude mcp add vestige /usr/local/bin/vestige-mcp -s user

export HTTPS_PROXY=your-proxy:port

curl http://localhost:3927/api/health
# Should return {"status":"healthy",...}

More troubleshooting ->

Contributing

Issues and PRs welcome. See CONTRIBUTING.md.

License

AGPL-3.0 — free to use, modify, and self-host. If you offer Vestige as a network service, you must open-source your modifications.

Built by @samvallad33
_{80,000+ lines of Rust · 30 cognitive modules · 130 years of memory research · one 22MB binary}