PAMbaseDocs
Build

Remember & recall.

PAMbase gives the user one persistent AI agent — a light identity plus a memory that follows them across every app. Memory is natural language: your app writes plain text; an LLM is what later reads it and decides what to do. This page is the whole read/write loop.

Prerequisites
A connection token from the connect flow, with memory:write:<scope> to write and memory:read:<scope> to read.

Write: remember()

Any app can remember something — no LLM required, because it's just text. Pass a string, or an object with a kind and a scope so other apps' agents can find it later.

typescript
// Shorthand: just text (defaults to kind "event", scope "general").
await pambase.remember("Prefers window seats; dislikes spicy food.");


// Better: categorize it so it's recallable across apps.
await pambase.remember({
  content: "Finished a 5km run in 28:10.",
  kind: "event",            // advisory hint; common kinds autocomplete, any string is fine
  scope: "fitness.running", // top-level must be a standard namespace; sub-scopes are free
  importance: 50,           // 0–100, optional
});


// Batch several at once (kind is free-form — use your own when none fit):
await pambase.remember([
  { content: "Saving for a house down payment.", scope: "finance" },
  { content: "Sister Maya lives in Toronto; they talk weekly.", kind: "relationship", scope: "family" },
  { content: "Draft the Q3 roadmap before the offsite.", kind: "task", scope: "note.work" },
]);
// → { accepted: true, memoryIds: ["mem_…", "mem_…"] }

Where memories go: kind × scope

A memory is a kind (an advisory structural hint) plus a scope (its domain). The two are not symmetric: kind is free-form — the common values just autocomplete and help the reading LLM — while scope is controlled. The scope's top-level namespace must come from the standard taxonomy (so a fitness app and a health app write to the same places and a wellness agent can recall across them); sub-scopes like fitness.running are free. An unknown top-level scope is rejected — use general if nothing fits.

FieldWhat it is
kindAn advisory hint — commonly fact, preference, event, relationship, emotion, goal, world, or summary, but free-form (e.g. task). Defaults to event.
scopeA domain from the taxonomy (e.g. health, finance, travel, learning) + an optional sub-scope.
contentThe natural-language text itself — what an LLM reads.

Read: recall()

To use memory, bring an LLM. recall() returns the memories most relevant to a query (hybrid semantic + keyword + recency + importance) as natural language — your agent interprets them and decides the next move.

typescript
const { memories } = await pambase.recall({ query: "what should they read next?", limit: 5 });
// memories[]: { id, type, scope, content, importance, occurredAt, sourceApp }


const system = "You are a reading companion. What's known about the user:\n" +
  memories.map((m) => "- " + m.content).join("\n");
const reply = await yourLLM.chat({ system, messages: [{ role: "user", content: "Recommend a book." }] });

No model of your own? Host the agent — PAMbase's chat is already grounded in this memory and records new memories from the conversation automatically. For a one-line portrait of the user, use the brief.

Permission filtering on writes

A memory whose scope isn't covered by one of your memory:write:* grants is silently dropped before it persists. The response's memoryIds reflects only what actually landed — it's the ground truth for what your write stored.

Next