SSM.js

JavaScript-native AI runtime — SSM execution + Transformer orchestration + online distillation + persistent agent memory.

SSM.js is a complete, self-contained AI runtime built directly on top of MambaCode.js. It includes the full session layer (previously @seanhogg/mambakit) as an internal layer, so you only need one package.

---

Overview

SSM.js is a JavaScript-native AI runtime that combines local SSM (State Space Model) inference with optional transformer bridge escalation, persistent semantic memory, and online distillation — all without leaving the browser or Node.js process.

The layered stack:

MambaCode.js  →  WebGPU kernels (WGSL, Mamba-1/2/3 SSM math)
SSM.js        →  Session layer + Runtime orchestration (this package)
                 ├── src/session/   MambaSession, tokenizer, persistence
                 ├── src/runtime/   SSMRuntime, routing
                 ├── src/memory/    MemoryStore
                 ├── src/agent/     SSMAgent
                 └── src/distillation/  DistillationEngine

Capability	SSM.js
Simple session API	✅
WebGPU execution	✅
SSM variants (1/2/3/hybrid)	✅
Transformer bridge	✅
Intelligent routing	✅
Online distillation	✅
Persistent semantic memory	✅
Agent workflows	✅

---

Installation

npm install @seanhogg/ssmjs
# or
pnpm add @seanhogg/ssmjs

@seanhogg/ssmjs includes the full session layer (previously @seanhogg/mambakit). @seanhogg/mambacode.js is a peer dependency — install it alongside:

npm install @seanhogg/ssmjs @seanhogg/mambacode.js

Node.js requirements

Node.js 18+ is required. Two additional shims are needed for Node.js:

npm install @webgpu/node fake-indexeddb

• @webgpu/node — Dawn-based WebGPU for Node.js; drives all WGSL compute kernels
• fake-indexeddb — in-memory IndexedDB compatible with the IDB spec; used by MemoryStore

---

Quick Start

Browser

import { SSM, AnthropicBridge, SSMAgent, MemoryStore } from '@seanhogg/ssmjs';

const runtime = await SSM.create({
  session: { modelSize: 'small', mambaVersion: 'mamba2' },
  bridge : new AnthropicBridge({ apiKey: 'sk-ant-...' }),
});

// Generate — routes to SSM or transformer automatically
const answer = await runtime.generate('What is a state space model?');

// Streaming — always SSM for low-latency output
for await (const token of runtime.stream('function fibonacci(')) {
  process.stdout.write(token);
}

// Fine-tune on your content
await runtime.adapt(myCodebase);
runtime.destroy();

Node.js

import { create as createGPU } from '@webgpu/node';
import { IDBFactory }          from 'fake-indexeddb';
import { SSM, MemoryStore, SSMAgent } from '@seanhogg/ssmjs';

const gpuAdapter = await createGPU().requestAdapter({ powerPreference: 'high-performance' });
const idbFactory = new IDBFactory();

const runtime = await SSM.create({
  session: {
    gpuAdapter,
    idbFactory,
    modelSize: 'small',
  },
});

const memory = new MemoryStore({ idbFactory });
const agent  = new SSMAgent({ runtime, memory });
await agent.init();   // loads persisted history if present

const reply = await agent.think('Explain this codebase');
console.log(reply);

await agent.destroy(); // persists history, releases GPU

---

Custom Tokenizers

By default, MambaSession uses the built-in Qwen2.5-Coder BPE tokenizer. You can override this by passing any object that satisfies the Tokenizer interface:

import type { Tokenizer } from '@seanhogg/ssmjs';

const myTokenizer: Tokenizer = {
  encode(text: string): number[]  { /* your encode implementation */ return []; },
  decode(tokens: number[]): string { /* your decode implementation */ return ''; },
  get vocabSize(): number          { return 32000; },
};

const runtime = await SSM.create({
  session: {
    tokenizer: myTokenizer,   // replaces BPETokenizer entirely
    modelSize: 'small',
  },
});

Use cases:

• HuggingFace Transformers.js tokenizer — wrap its encode/decode in the interface
• Unit testing — a stub tokenizer that maps words to sequential IDs, no network needed
• Domain-specific vocabularies — medical, legal, multilingual tokenizers

---

Memory System

MemoryStore is a persistent, TTL-aware, tagged key-value fact store backed by IndexedDB.

Basic usage

import { MemoryStore } from '@seanhogg/ssmjs';

const memory = new MemoryStore({
  dbName      : 'my-app',
  defaultTtlMs: 7 * 24 * 60 * 60 * 1000,  // 7-day default TTL
  idbFactory,  // Node.js only
});

// Store facts
await memory.remember('author', 'Sean Hogg');
await memory.remember('stack', 'React + TypeScript', {
  tags      : ['tech', 'project'],
  importance: 0.8,
  ttlMs     : 30 * 24 * 60 * 60 * 1000,  // 30 days
});

// Retrieve
const entry = await memory.recall('author');

// All non-expired facts, newest first
const all = await memory.recallAll();

// N most recent non-expired facts
const recent = await memory.recallRecent(10);

// Filter by tag
const techFacts = await memory.recallByTag('tech');

// Semantic similarity search (Jaccard word-overlap; SSM embeddings in future)
const similar = await memory.recallSimilar('who built this?', 5, runtime);

// Purge expired entries from storage
const deletedCount = await memory.purgeExpired();

Cross-session memory merge

// Export all non-expired facts from sessionA
const exported = await memoryA.exportAll();

// Import into sessionB
await memoryB.importAll(exported, 'merge');
// 'merge'     — only overwrites if incoming entry is newer
// 'overwrite' — writes all entries unconditionally

Weight persistence

await memory.saveWeights(runtime);      // saves model weights to IndexedDB
const loaded = await memory.loadWeights(runtime);  // false if no checkpoint found

MemoryEntry schema

interface MemoryEntry {
  key        : string;
  content    : string;
  timestamp  : number;
  ttlMs?     : number;       // optional TTL; entry filtered after timestamp + ttlMs
  type?      : FactType;     // 'text' | 'json' | 'number' | 'boolean'
  tags?      : string[];     // for grouping/filtering
  importance?: number;       // 0–1, default 0.5; higher facts appear first in prompts
}

---

Inference Routing

InferenceRouter decides whether each request goes to the local SSM or the transformer bridge. It is built into SSMRuntime — you don't need to instantiate it directly.

Routing strategies

const runtime = await SSM.create({
  session: { modelSize: 'nano' },
  bridge : claude,
  routingStrategy   : 'auto',   // 'auto' | 'ssm' | 'transformer'
  longInputThreshold: 1200,     // chars before preferring transformer (default: 1200)
  perplexityThreshold: 80,      // SSM perplexity cutoff (default: 80)
});

Auto-routing heuristics (cheapest first):

1. Complexity patterns — "step by step", "analyze", "compare and contrast" → transformer
2. Input length — over threshold → transformer
3. SSM perplexity — async probe; high perplexity = novel topic → transformer

RoutingDecision type

route() now returns a structured RoutingDecision object:

interface RoutingDecision {
  target    : 'ssm' | 'transformer';
  reason    : 'strategy' | 'complexity' | 'length' | 'perplexity' | 'no_bridge';
  confidence: number;    // 0–1
  details?  : string;    // human-readable explanation
}

Routing audit log

Every routing decision is appended to an in-memory audit log (last 500 entries):

const log = runtime.getRoutingAuditLog();
// log: RoutingAuditEntry[]
// { timestamp, inputLength, decision: RoutingDecision, durationMs }

---

Distillation

Teach the local SSM using a transformer teacher — runs entirely in the browser or Node.js.

import { DistillationEngine } from '@seanhogg/ssmjs';

const distiller = new DistillationEngine(runtime, claude);

// Single pass: claude generates → SSM adapts on output
const result = await distiller.distill('Explain WebGPU compute shaders', {
  adapt      : { wsla: true, epochs: 3 },
  qualityGate: {
    minLength    : 50,    // skip if teacher output < 50 chars
    maxPerplexity: 15,    // skip if SSM perplexity already < 15 (already learned)
  },
});

console.log('skipped:', result.skipped, result.skipReason);
console.log('loss:',    result.adaptResult.losses.at(-1));

// Batch distillation
const batch = await distiller.distillBatch([
  'What is a Mamba block?',
  'Explain WSLA adaptation.',
], { adapt: { wsla: true, epochs: 5 } });

console.log(`${batch.totalEpochs} epochs in ${batch.totalMs}ms`);

Quality gates

Gate option	Description
minLength	Skip if teacher output is shorter than N characters (low-quality response)
maxPerplexity	Skip if SSM perplexity on teacher output is already below threshold (already learned)

Distillation log

const log = distiller.getLog();
// log: DistillationLog[]
// { timestamp, input, teacherOutputLength, skipped, skipReason?, finalLoss?, epochs }

The log is bounded to the last 200 entries.

---

SSMAgent

High-level orchestration: conversation history, routing, memory injection, and lifecycle.

import { SSMAgent, MemoryStore } from '@seanhogg/ssmjs';

const memory = new MemoryStore();
const agent  = new SSMAgent({
  runtime        : runtime,
  memory,
  systemPrompt   : 'You are a senior TypeScript engineer.',
  maxHistoryTurns: 20,
  persistHistory : true,  // saves/loads history via memory on destroy/init
});

// Load persisted history from a prior session
await agent.init();

// Store project context
await agent.remember('stack', 'React 18, TypeScript 5, Vite');

// Multi-turn conversation — facts with highest importance appear first in context
const reply1 = await agent.think('What stack should I use?');
const reply2 = await agent.think('How do I handle concurrent edits?');

// Streaming
for await (const token of agent.thinkStream('Show me a WebSocket hook')) {
  process.stdout.write(token);
}

// Teach the agent from content
await agent.learn(myCodebase, { wsla: true, epochs: 3 });

console.log(agent.turnCount);  // 2

// Persists history to memory, then destroys runtime
await agent.destroy();

History persistence

When persistHistory: true (default):

• On agent.init(): loads __history__ from the MemoryStore and restores conversation turns.
• On agent.destroy(): serialises _history to JSON and writes it under __history__.

This enables multi-session continuity without external state management.

Fact injection order

Facts retrieved from MemoryStore are sorted by importance descending before being injected into the prompt. Higher-importance facts appear first, giving the model the most relevant context regardless of insertion order.

---

Migration from MambaKit

@seanhogg/mambakit has been consolidated into this package. MambaSession and all related types are now exported directly from @seanhogg/ssmjs.

Before:

npm install @seanhogg/mambakit @seanhogg/ssmjs

import { MambaSession } from '@seanhogg/mambakit';
import type { MambaSessionOptions, Tokenizer } from '@seanhogg/mambakit';

After:

npm install @seanhogg/ssmjs

import { MambaSession, SessionError } from '@seanhogg/ssmjs';
import type { MambaSessionOptions, Tokenizer } from '@seanhogg/ssmjs';

All types are re-exported unchanged — MambaSessionOptions, CompleteOptions, AdaptOptions, AdaptResult, SaveOptions, LoadOptions, StorageTarget, CreateCallbacks, LayerSchedulePreset, MODEL_PRESETS, GpuMode, and Tokenizer. No logic changes are required, only the import path.

---

CoderClaw Integration

SSM.js serves as the hippocampus layer of CoderClaw's gateway — a persistent semantic memory and local inference engine running alongside the frontier LLM (Claude/GPT) cortex.

The SsmMemoryService class in CoderClaw's src/infra/ssm-memory-service.ts wraps an SSMRuntime + SSMAgent + MemoryStore triplet:

CoderClaw gateway
├── server-startup.ts       ← initSsmMemoryService() on boot
├── infra/knowledge-loop.ts ← remember() + learn() on every agent run
├── infra/ssm-memory-service.ts  ← SsmMemoryService wrapper
└── coderclaw/orchestrator.ts    ← recallSimilar() injected into task prompts

Data flow:

1. Agent run completes → KnowledgeLoopService derives activity summary
2. Summary is stored in .coderClaw/memory/YYYY-MM-DD.md (markdown log)
3. Summary is also passed to ssmSvc.remember() (tagged + importance-weighted)
4. Summary is passed to ssmSvc.learn() → WSLA fine-tuning adapts the SSM
5. On next workflow task, recallSimilar(taskDescription, 5) injects relevant memories into the prompt as a [Memory Context] block

GPU init is optional: if @webgpu/node is unavailable, the service starts in memory-only mode (gpuAvailable: false) and SSM inference is skipped. The gateway never crashes due to a missing GPU.

---

Phase Roadmap

Phase 1 — Foundations

• Session layer: Tokenizer interface + pluggable injection via MambaSessionOptions.tokenizer
• MemoryStore: TTL (ttlMs), defaultTtlMs, purgeExpired(), recallRecent(n)
• MemoryStore: FactType, tags, importance fields on MemoryEntry
• MemoryStore: updated remember() accepting RememberOptions
• InferenceRouter: route() now returns RoutingDecision object with target, reason, confidence, details
• SSMAgent: persistHistory option; init() loads __history__; destroy() saves it
• SSMAgent: fact injection sorted by importance descending

Phase 2 — Semantic Memory

• MemoryStore.recallSimilar(query, topK, runtime) — Jaccard similarity; SSM embedding-based search in future
• MemoryStore.recallByTag(tag) — tag-based filtering
• MemoryStore.exportAll() / importAll(entries, strategy) — cross-session merge

Phase 3 — CoderClaw Integration

• SsmMemoryService in src/infra/ssm-memory-service.ts — singleton gateway service
• server-startup.ts: initSsmMemoryService() on boot; non-fatal GPU fallback
• KnowledgeLoopService: remember() + learn() on every agent run completion
• AgentOrchestrator: recallSimilar() injected as [Memory Context] before task dispatch

Phase 4 — Feedback Loop

• DistillationEngine: qualityGate option (minLength, maxPerplexity)
• DistillResult: skipped + skipReason fields
• DistillationEngine.getLog() — bounded in-memory DistillationLog[]
• InferenceRouter: RoutingAuditEntry + getAuditLog() — bounded in-memory log
• SSMRuntime.getRoutingAuditLog() — delegates to router
• SSMRuntime.getDistillationLog() — stub; returns empty array (inline engine future work)

---

API Reference

SSM.create(opts) / SSMRuntime.create(opts)

Option	Type	Default	Description
session	MambaSessionOptions	required	Forwarded to MambaSession.create()
bridge	TransformerBridge	—	Transformer backend for routing/distillation
routingStrategy	'auto'|'ssm'|'transformer'	'auto'	Routing strategy
longInputThreshold	number	1200	Chars before auto-routing prefers transformer
perplexityThreshold	number	80	SSM perplexity cutoff
callbacks	CreateCallbacks	—	Progress callbacks

runtime.generate(input, opts?)

Generates a full response. Routes to SSM or transformer per strategy. Returns Promise<string>.

runtime.stream(input, opts?)

AsyncIterable<string> — always uses SSM path for consistent latency.

runtime.streamHybrid(input, opts?)

AsyncIterable<string> — routes like generate(), streams via bridge if available.

runtime.adapt(data, opts?)

Pass-through to session.adapt(). Returns AdaptResult.

runtime.evaluate(text)

Returns SSM perplexity. Used internally by auto-routing.

runtime.getRoutingAuditLog()

Returns RoutingAuditEntry[] — last 500 routing decisions with timing.

runtime.getDistillationLog()

Returns DistillationLog[] — last 200 distillation runs (stub; use distiller.getLog() directly).

runtime.save(opts?) / runtime.load(opts?)

Weight persistence pass-throughs to MambaSession.

runtime.destroy()

Releases GPU device and all buffers.

---

Error Handling

import { SSMError, SessionError } from '@seanhogg/ssmjs';

try {
  const runtime = await SSM.create({ session: { modelSize: 'nano' } });
  await runtime.generate('hello');
} catch (err) {
  if (err instanceof SSMError) {
    // Runtime-level error (bridge, distillation, memory)
    console.error(err.code);  // 'BRIDGE_REQUEST_FAILED' | 'RUNTIME_DESTROYED' | ...
  }
  if (err instanceof SessionError) {
    // Session-level error (GPU init, tokenizer, checkpoint)
    console.error(err.code);  // 'GPU_UNAVAILABLE' | 'TOKENIZER_LOAD_FAILED' | ...
  }
}

---

File Structure

src/
├── index.ts                          ← package entry + SSM namespace
├── session/                          ← session layer (absorbed from @seanhogg/mambakit)
│   ├── session.ts                    ← MambaSession.create() — GPU, tokenizer, model, persistence
│   ├── tokenizer.ts                  ← Tokenizer interface (pluggable)
│   ├── presets.ts                    ← MODEL_PRESETS + layer schedule resolution
│   ├── persistence.ts                ← IndexedDB / download / File System API helpers
│   ├── streaming.ts                  ← AsyncIterable token streaming
│   ├── errors.ts                     ← SessionError typed error class
│   └── index.ts                      ← barrel export
├── runtime/
│   └── SSMRuntime.ts                 ← core runtime, owns MambaSession
├── bridges/
│   ├── TransformerBridge.ts          ← interface
│   ├── OpenAIBridge.ts               ← OpenAI chat completions
│   ├── AnthropicBridge.ts            ← Anthropic Messages API
│   └── FetchBridge.ts                ← generic OpenAI-compatible endpoint
├── router/
│   └── InferenceRouter.ts            ← SSM ↔ transformer routing + audit log
├── memory/
│   └── MemoryStore.ts                ← IndexedDB fact store: TTL, tags, importance, export/import
├── distillation/
│   └── DistillationEngine.ts         ← online teacher→student distillation + quality gates
├── agent/
│   └── SSMAgent.ts                   ← orchestration: history persistence + fact injection
└── errors/
    └── SSMError.ts                   ← typed error class

---

Professional Platform

SSM.js patterns are the architectural foundation of Builderforce.ai's Agent Runtime.

SSM.js concept	Builderforce.ai equivalent
SSMRuntime	AgentRuntime (browser-native, ties to IDE project)
DistillationEngine	LLM-assisted dataset generation → in-browser LoRA training
MemoryStore	IndexedDB MambaAgentState + AgentPackage embedding
SSMAgent	Published workforce agent (Workforce Registry)
TransformerBridge	Cloudflare Workers AI / OpenRouter fallback

---

License

MIT