@juspay/neurolink/dist/memory/condensationPrompt.js

The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002

/**
 * Placeholders `@juspay/hippocampus` substitutes into a condensation prompt
 * (`replaceAll` in `Hippocampus.add()`). None is required by Hippocampus
 * itself — a missing one simply stays literal — which is exactly why the
 * two structural ones must be checked here: a prompt without
 * `{{NEW_CONTENT}}` produces a memory that never grows, and one without
 * `{{OLD_MEMORY}}` overwrites the whole summary on every turn. Both fail
 * silently at the LLM boundary.
 */
const CONDENSATION_PLACEHOLDERS = [
    "OLD_MEMORY",
    "NEW_CONTENT",
    "MAX_WORDS",
];
const FATAL_PLACEHOLDERS = new Set([
    "OLD_MEMORY",
    "NEW_CONTENT",
]);
/**
 * Check a condensation prompt template for the placeholders Hippocampus
 * expects. Pure — never logs or throws. `valid` is false only when a
 * structural placeholder (`OLD_MEMORY` / `NEW_CONTENT`) is missing;
 * a missing `MAX_WORDS` is reported in `missing` but leaves `valid` true.
 */
export function validateCondensationPrompt(prompt) {
    const missing = CONDENSATION_PLACEHOLDERS.filter((name) => !prompt.includes(`{{${name}}}`));
    const fatal = missing.filter((name) => FATAL_PLACEHOLDERS.has(name));
    return { valid: fatal.length === 0, missing, fatal };
}