feat(sona): Save/load learned state to disk for persistence across restarts

[dmoellenbeck]

Feature Request

SONA's learned weights (MicroLoRA, BaseLoRA, EWC++ Fisher matrix, ReasoningBank patterns) are in-memory only. On process restart, all learning is lost and must be re-acquired by replaying trajectories.

Motivation

For long-running applications (coaching systems, production chatbots, persistent agents), SONA needs to survive process restarts without losing learned adaptations. Currently, the only workaround is to re-feed all trajectories on every startup, which:

• Requires >= 100 trajectories before forceLearn() activates
• Takes ~30s for 200 trajectories via MCP stdio
• Loses any patterns that were learned during the previous session

Proposed API

Rust:

// Save learned state (LoRA weights, EWC Fisher, patterns)
engine.save_state("path/to/sona-state.bin")?;

// Load from previous state
let engine = SonaEngine::load_state("path/to/sona-state.bin")?;

Node.js (NAPI):

engine.saveState("path/to/sona-state.json");
const engine = SonaEngine.loadState("path/to/sona-state.json");

What should be persisted

Component	Estimated Size (256-dim)
MicroLoRA weights (rank-2)	~2 KB
BaseLoRA weights (rank-8)	~16 KB
EWC++ Fisher information matrix	~256 KB
ReasoningBank cluster centroids	~40 KB
Pattern metadata	~5 KB
Total	~320 KB

Current workaround

We store trajectories in a JSON file (intelligence.json) and replay them on startup. The @ruvector/cli hooks session-start command initializes the intelligence layer, but SONA's neural weights start fresh each time.

Additional context

The IntelligenceEngine in ruvector npm already persists Q-tables, memories, and trajectories to intelligence.json. SONA state could be added to this same file, or stored separately as a binary blob for efficiency.

Related: hooks_force_learn requires 100 trajectories minimum even with force=true (#273 covers the stats counter bug). State persistence would eliminate the need for trajectory replay entirely.

[ruvnet]

Partial fix in PR #284 (merged).

What's done:

• Added serialize_state() to LoopCoordinator — returns JSON with reasoning bank patterns, EWC task count, enabled flags
• Added saveState() NAPI binding
• Added saveState() to TypeScript wrapper

const state = engine.saveState();  // JSON string
fs.writeFileSync('sona-state.json', state);
// On restart: load and re-hydrate (loadState not yet implemented)

What's remaining:

• loadState() / deserialize_state() to restore from saved JSON on startup
• Integration with brain server to auto-persist to Firestore on each training cycle
• Auto-load on Cloud Run cold start

The save path is in place; the load path needs a follow-up PR.

Files: crates/sona/src/loops/coordinator.rs, crates/sona/src/engine.rs, crates/sona/src/napi.rs, npm/packages/ruvector/src/core/sona-wrapper.ts

[dmoellenbeck]

I'm actively using SONA for an adaptive coaching system (130+ real-world
trajectories, daily check-ins via MCP server).

Current state in ruvector@0.2.18:

• saveState() and loadState() are defined in the JS wrapper ✓
• But the native binary (@ruvector/sona@0.1.5) does not export these methods ✗
• this._native.saveState is not a function at runtime

Filed #301 — the @ruvector/sona binary needs to be rebuilt with the current
napi.rs code which has both save_state() and load_state() implementations.

My workaround: scripts/sona-train.mjs replays ~132 trajectories from
SQLite on every MCP server start (~3s). Works but loses any patterns learned
during the session. Once saveState()/loadState() ship, we'll persist to
~/.claude-coach/ruvector/sona-state.json and eliminate the replay overhead.

Happy to help test the rebuilt binary when it's available.