Skip to main content

The memory base

The memory base is where Agivar keeps, long-term, "the things you taught it". Teach Mode writes to it; Task Mode reads from it.

Structure: platform / topic​

Every memory lives under a platform/topic path:

  • platform — the name of an app, website, or system, e.g. bilibili, vscode, company-expense-system. Cross-platform know-how, or anything with no clear owner, goes under general.
  • topic — a specific operation, workflow, rule, or piece of know-how, e.g. coins, submit-invoice, tool-call-preference.

Examples:

bilibili/coins
bilibili/triple-action
company-expense-system/submit-invoice
general/expense-rules
general/tool-call-preference

What a memory entry looks like​

Two kinds, by content:

Operating workflows (the most common)​

When you teach the workflow of an app/website/system, the entry contains:

  1. Overview — the goal, applicable scenario, and end result of the operation.
  2. Initial state — what state things should be in before executing (browser open? logged in? which page? what key elements should be visible?).
  3. Steps — split into several abstract steps; each has its concrete operations underneath, ending with a "stage result". On-screen targets are written unambiguously.
  4. Related demo recording (if there is one) — recording ID, summary, and key clips relevant to this entry.

(See the full example at Teach by recording → Step 6.)

Rules / preferences / concepts​

E.g. "expense invoices no more than ¥500 don't need an attached note", "I like tool calls merged into one actions array", "what we call a 'frozen order' means…" — organized freely by content, but kept clearly structured and reusable.

How the AI maintains it​

  • Look at the index before deciding new vs. update — the Teach/Task agent has the current memory-base index on hand from the start (one line per entry: platform/topic + a one-line summary), so it knows what entries exist. To change one, it reads the original text first, then makes a precise local edit — rather than creating a duplicate.
  • Organize, don't copy verbatim — it turns what you said into something clear, executable, and reusable.
  • Latest wins on conflict — when old and new information conflict, your latest explicit statement wins.
  • When unsure, mark it or ask — it won't invent steps you didn't teach.
  • Not everything gets stored — casual chat, one-off instructions, and information with no reuse value are not filed.

Viewing / editing​

  • View: in a Teach-mode (or Task-mode) conversation, just ask it to "list the memory entries related to X", or "read the bilibili/coins entry to me".
  • Edit / correct: in a Teach-mode conversation, point out what's wrong or what to add, and it updates that entry — see Teach by chat → Correcting or supplementing an existing memory.
  • Delete: ask it to delete a memory you no longer need.

How Task Mode uses it​

Before running a GUI task, Task Mode consults the memory base: as long as an entry might have something useful for the current task, it reads it; entirely unrelated ones are skipped. If an entry it reads references a demo recording (like Recording ID: 00046), and that recording is useful for the concrete steps, it asks targeted questions about details in it.

So — the more you teach in Teach Mode, and the more accurately, the easier and more reliable Task Mode becomes.

Multiple devices / accounts​

The memory base lives in the cloud, tied to your account. Sign in with the same account on another computer and it's still there; different accounts have independent memory bases. (A recording's raw frames stay local on the computer that recorded it; the cloud keeps the processed keyframes and text results.)

Next​