DEV Community

t49qnsx7qt-kpanks
t49qnsx7qt-kpanks

Posted on

task memory is what makes agents stop redoing yesterday's work

task manager mcp shipped on hn today - opensource, supervisor mcp + worker workspaces. clean architecture. the missing piece for most agent teams is the same one this build hints at - memory across runs.

what most agents do today

run the task. log the completion. forget it.

next run, a different agent does the same task again because nothing told it the previous one already shipped.

what task memory does

persist the task graph - what was attempted, what succeeded, what failed, what the operator approved. next agent reads the graph before it starts.

result - 30-50% reduction in redundant tool calls in the agent benchmarks i've run.

the two-line integration

import { mnemopay } from '@mnemopay/sdk';
await mnemopay.remember({ task_id, status, context, parent_workflow });
const prior = await mnemopay.recall({ task_id });
Enter fullscreen mode Exit fullscreen mode

drops into any mcp supervisor pattern. the worker queries recall before it starts. the supervisor stores remember after it lands.

why mnemopay specifically

two reasons:

  • the memory persists across mcp servers, agent frameworks, and platforms (your mcp supervisor today might run on agentcore tomorrow)
  • every memory write emits a reputation signal - agents that complete tasks reliably build score over time

what the bizsuite plugin does

wraps the mcp tool list with mnemopay.remember calls. ships as a single npm install. agent forgets nothing.

Top comments (0)