a research agent that forgets is a search engine with extra steps

almanac mcp shipped on hn - turns claude code into a deep research agent. nice build. the api is clean and the documentation is tight.

the 80/20 question for any research agent - does it persist what it found.

the failure mode

day 1 - agent researches the iea 2025 report on data center demand. summarizes. ships.

day 30 - same operator, same question, fresh session. agent re-fetches the same pdf, re-summarizes, ships the same answer.

that's a search engine with extra steps. the value of an agent vs google is supposed to be that it remembers.

what persistent research looks like

import { mnemopay } from '@mnemopay/sdk';

const prior = await mnemopay.recall({ topic, depth: 'cited_sources' });
if (prior?.confidence > 0.8) return prior.summary;

const fresh = await almanac.research({ topic });
await mnemopay.remember({ topic, summary: fresh, sources, ts });
return fresh;

four lines. the agent now skips the redundant research and incrementally extends what it already knows.

why this matters for the cost line

a deep research run on claude opus 4 burns 80k-200k tokens. memoizing the result saves $0.40-$1.50 per repeated query. on an agent doing 1000 research tasks a month, that's $400-1500 saved by not forgetting.

almanac is the engine. mnemopay is the cache.