@zanii/memory

Provable agent memory. A memory.write receipt hash-chains what an agent remembered and when — so an auditor can answer "why did it decide that?" and prove the memory wasn't edited after the fact.

Each entry commits its content with a salt (the raw memory never leaves your system) and links to the previous entry's entry_hash — a tamper-evident chain. The entry_hash is an RFC 8785 canonical hash (@zanii/core), so a chain written in TypeScript verifies byte-identically in Python.

npm install @zanii/memory @zanii/core

Usage

import { appendMemory, verifyMemoryChain, verifyContent } from '@zanii/memory';

// write a chain — appendMemory auto-links prev + increments seq:
const a = appendMemory(null, { content: 'user asked for a refund', kind: 'observation', ts });
const b = appendMemory(a,    { content: 'policy allows refund within 30 days', kind: 'fact', ts });
const c = appendMemory(b,    { content: 'issued the refund', kind: 'decision', ts });

await agent.record(c.target, c.payload);   // keep c.salt to prove the content later

// audit: the chain is intact (nothing inserted, deleted, or edited):
verifyMemoryChain([a, b, c]).ok;           // true

// prove what a specific entry actually remembered:
verifyContent('issued the refund', c.salt, c.payload.content_commitment); // true

buildMemoryWrite(input) builds a single entry when you manage prev/seq yourself. verifyMemory(payload) checks one entry's entry_hash is self-consistent (catches an edited entry); verifyMemoryChain(entries) also checks the links and that seq increments.

Changelog

• 0.1.0 — initial release: commitContent / verifyContent, buildMemoryWrite, appendMemory, verifyMemory, verifyMemoryChain.

License

Apache-2.0.