@halo-sdk/memory

Cache-aware long-term memory for Halo AI SDK.

Memory lives in a single cache segment placed before the volatile history tail, so it sits inside the prefix cache and stays byte-stable — and the cache stays warm — until an explicit write. remember() is append-only (keeps the existing cached prefix valid); forget() / clear() rebuild the block as a deliberate, attributable cache miss.

Usage

import { Halo } from "@halo-sdk/core";
import { MemoryStore } from "@halo-sdk/memory";

const memory = new MemoryStore();
memory.remember("The user prefers TypeScript.");
memory.remember("The user is in UTC+8.");

const agent = halo.agent({
  /* ... */
});
agent.setContextSegments([memory.segment]);

// later
memory.recall(); // ["The user prefers TypeScript.", "The user is in UTC+8."]

Each entry is its own message, so appends extend the block without touching earlier entries — the prefix up to the previous tail stays cached.