AI agent memory infrastructure
that saves tokens and keeps context

// .env — point base URL at TES, add your token
ANTHROPIC_API_KEY=sk-ant-...
ANTHROPIC_BASE_URL=https://llm.api.pentatonic.com
TES_API_KEY=tes_<clientId>_<random>

// Wire it on your client (one extra line)
import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic({
  defaultHeaders: { "X-TES-Token": process.env.TES_API_KEY },
});
await client.messages.create({ model, messages });

// What TES does on the wire (you don't write this)
const context = await retrieve({
  session: req.session,
  files: req.referenced_files,
  memory: req.user_memory,
});

req.messages = [preamble(context), ...req.messages];
// e.g. on a code-lookup turn from the published benchmark:
// → 10,050 tokens of tool/file work avoided
// → 17 follow-up tool calls eliminated

// Response is identical to direct upstream call
{
  id: "msg_01...",
  model: "claude-sonnet-4",
  content: [{ type: "text", text: "..." }],
  usage: { input_tokens: 450 /* was 10500 OFF */ },
  // overall benchmark median: 27.2% reduction.
  // code-lookup category median: 95.6%. See /benchmarks.
}