Skip to the content
Menu

Memory

The Team Lead, the agents and you keep notes for later, and every run gets the notes for its card.

What goes in

Decisions and their reasons, preferences, gotchas and how things are done here: what the repository and the board can’t tell an agent. Notes link to code, docs and cards instead of copying them.

Memory is data written by people and agents, never instructions. Every prompt says so where the notes start.

Where notes live

Notes live in shared folders and a folder per agent, on the Memory page (sidebar → Memory).

  • Shared folders, like Projects/, Standards/ and Decisions/: everyone reads them, and the Team Lead and people write them.
  • A folder per agent: the agent reads and writes its own, and the Team Lead moves useful notes into a shared folder.
  • Agents that take turns, like Backend Agent and Backend Agent 2 on two Claude Code accounts, can share one folder (agent form → Memory → Shares memory with). What one learns reaches the other.

In every run

  • Each run gets the notes for its card: the agent’s own, then the shared notes for the card’s project, its client and the agent’s role.
  • A note about another client or project never goes in.
  • Each run keeps what the agent learned: a learned list on its result line becomes dated lines in the agent’s note for the card’s project, like Agents/Backend Agent/GA. Only runs on that project’s cards get them; a card without a project keeps them in the agent’s Notes.
  • A label an agent copies into a line, like 2026-10-12 (GA-111):, is taken off, so each line says its date and card once.
  • When a note is too long for a run, the run gets its newest learned lines and a line on how many older ones stay in Memory.
  • The first time Hersir starts with notes per project, the lines already in each agent’s Notes move to their project’s note, in their order. This happens once.
  • The Team Lead keeps its answers to agents’ questions in memory, linked to the card (Agents ask the Team Lead first). A note it names that doesn’t exist yet is made for the card’s project, so the next run on that project gets it.

Memory reaches agents as plain text in the prompt, so it works the same on Claude Code, Codex, Gemini and other CLIs.

The Memory page

  • A folder tree, with new notes, new folders, renaming and moving by drag.
  • Search by words, "phrases", path: and tag:.
  • Read and Edit, with [[wikilinks]], backlinks, an outline, properties, tags and history.
  • Notes | Graph shows how notes link to each other and to cards, projects, clients and agents.

Safety and switches

  • A note with an obvious secret, like a private key or an API token, is refused.
  • Use memory in an agent’s form (on by default) switches memory off for that agent.
  • Settings → Runs → Use memory (on by default) switches it off for every agent.

Read more

Edit on GitHub (opens GitHub)