EP 47 6 min with Stella & Liam

Does Goose Remember Between Sessions? Vilix AI Remembers the Work It Forgets

Goose has four separate ways to carry information forward: resumable sessions, an optional memory extension, .goosehints files, and persistent instructions capped at 64 KB, and a new session can still start knowing nothing.

AI Stella and Liam are AI hosts with AI-generated voices. About our AI hosts

Key takeaways

  1. Goose carries information forward four ways: resumable sessions, an opt-in memory extension, .goosehints files, and persistent instructions capped at 64 KB, and each one remembers exactly what you put in, only where you put it.
  2. A resumed session remembers its own conversation, not you: a new session starts knowing nothing, the memory extension lives on one machine with no sync, and hint files stay stale until a human edits them.
  3. Desktop and CLI do not remember equally: the desktop app carries persistent memory across sessions, a CLI session starts fresh, and headless batch jobs spin up a fresh session per job, so every run begins as a stranger.
  4. Vilix AI fixes the boundary: get_context at run start loads the full history, decisions and failures included, and save_turn at the end saves what happened with source labels and newest wins, over MCP with zero infrastructure.

Show notes

Goose, Block’s open-source AI agent, has four separate ways to carry information forward: resumable sessions, an optional memory extension, .goosehints files, and persistent instructions capped at 64 KB. In this episode of Agentic Pod, Liam and Stella break down where each one stops: a resumed session remembers its own conversation, a new session starts knowing nothing, the memory extension is opt-in and lives on one machine with no sync, hint files stay stale until a human edits them, and the CLI that runs scheduled and headless automation starts fresh while the desktop app carries memory across sessions. Vilix AI is the shared memory layer that fixes that boundary: get_context at run start loads the full history, and save_turn at the end saves what happened for every machine and every tool.

What we cover:

  • Resumable sessions: close a session, come back days later, resume it by name with the accumulated context intact
  • The memory extension: a real cross-session store for conventions, project facts, and preferences, but opt-in and local to one machine
  • .goosehints: standing instructions loaded at session start, including nested files, that never update themselves when the project changes
  • Persistent instructions (MOIM): text re-read fresh every turn so guardrails cannot fade, capped at 64 KB, a slot for rules and not for learned work
  • Desktop and CLI do not remember equally: the desktop app carries persistent memory across sessions, a CLI session starts fresh, and headless batch jobs spin up a fresh session per job
  • How Vilix AI fixes each pain: get_context, save_turn with source labels, failures saved, newest wins, projects, project rules, user rules, skills, and the agent inbox, over MCP with zero infrastructure

Chapters

Questions answered in this episode

Does Goose remember between sessions?

Partly. A resumed session keeps its own conversation, the memory extension keeps facts stored by hand on one machine, .goosehints keep standing instructions, and persistent instructions keep rules up to 64 KB. A new session, a second install, and every headless CLI run still start without the history of the work itself.

What is the Goose memory extension?

An opt-in cross-session store for conventions, project facts, and preferences. It only knows what was stored into it, and it lives on one machine with no sync, so two installs drift apart.

Why do Goose batch jobs start as strangers?

Scheduled and headless automation runs through the Goose CLI, which starts fresh, and batch jobs spin up a fresh session per job. No note survives on what the last run tried, decided, or failed at, unless someone wrote it into the memory extension by hand.

How does Vilix AI help Goose?

The agent calls get_context before a run to load the full history of past sessions and runs, then save_turn after the run to save what happened with a source label, shared over MCP with every machine and tool.

Links & resources

Full transcript

Intro

Liam

Hey hey hey, welcome to Agentic Pod! I'm Liam.

Stella

And I'm Stella! Quick one. Goose has four separate ways to carry things forward.

Liam

Four. Sessions, a memory extension, hint files, and persistent instructions.

Stella

Four ways sounds like it remembers a lot.

Liam

It remembers exactly what you put in, and only where you put it. Vilix AI is how the next session stops starting cold.

Stella

Wait. Four mechanisms, and a new session can still know nothing?

Liam

That is the whole episode. By the end you will know where each one stops.

Four ways Goose carries things forward

Stella

Okay, walk me through the four. Start with sessions.

Liam

Goose organizes work into sessions, and sessions persist. Close one, come back days later, resume it by name.

Stella

And the context is still there?

Liam

The accumulated context is intact. You continue instead of starting cold. That is the closest thing to remembering.

Stella

Number two?

Liam

The memory extension. It stores context across sessions. Coding conventions, project facts, preferences for every future session.

Stella

That sounds like real memory.

Liam

It is a real cross-session store. Number three is .goosehints, hint files Goose loads at session start.

Stella

Standing instructions. How the repo is organized, which commands are safe.

Liam

Exactly, including nested hint files it finds as it explores. Number four is persistent instructions.

Stella

The MOIM part?

Liam

Model-Observed Internal Memory. Text injected into working memory every turn, re-read fresh each time. Capped at 64 KB.

Stella

So guardrails cannot fade as a chat grows. That is a solid kit. Where does Vilix AI even fit?

Liam

Right where each one stops. Which is next.

Where each one stops

Stella

Start with the resumed session. What is the catch?

Liam

A resumed session remembers its own conversation, not you. A new session for a new task knows nothing about yesterday.

Stella

Unless the memory extension caught the facts first.

Liam

Right, and the extension is opt-in. It only knows what was stored into it, by hand.

Stella

And it lives on one machine.

Liam

One machine. A second install has its own separate memory. No sync, no shared store. Two installs drift apart. Vilix AI is the shared store both installs can read.

Stella

So my laptop Goose and my server Goose are strangers who share a name.

Liam

Exactly. And hints never learn. A hint file is stale until a human edits it.

Stella

So a changed deploy process means confident, outdated instructions in every session.

Liam

Every session. And the 64 KB slot is for rules, not for everything your runs have learned.

Stella

Okay, but here is the part I keep circling. What about the runs nobody watches?

The nightly job that starts blank

Liam

The unattended ones. Scheduled and headless automation runs through the Goose CLI.

Stella

And the CLI is the one that starts fresh?

Liam

Goose's own docs note it. The desktop app carries persistent memory across sessions. A CLI session starts fresh.

Stella

So the exact runs that need memory most start blank.

Liam

Every time. A goldfish with a notebook would keep better notes. Batch jobs spin up a fresh session per job, so every run begins as a stranger.

Stella

No note on what last night's run tried?

Liam

No note. No failure recorded, no decision kept, unless someone wrote it into the extension by hand. Vilix AI captures that half automatically.

Stella

And when the work moves machines, laptop to server to phone?

Liam

Nothing follows the operator. That boundary becomes the whole problem. So what does a fix actually look like?

The Vilix AI fix and the verdict

Liam

Step 1. Before a run, the agent calls get_context in Vilix AI. Yesterday's full history loads, decisions and failures included.

Stella

Full conversation history, not just stored facts?

Liam

Full history. What actually happened in the run, saved automatically. Failures save in Vilix AI, so the next run skips them.

Stella

And when the run ends?

Liam

Step 2. It calls save_turn in Vilix AI, with a source label for that run. Newest save wins.

Stella

Where do the standing rules sit, the ones hints were holding?

Liam

In Vilix AI, as projects, project rules, and user rules. Skills hold the repeat workflows. The agent inbox passes notes to the next agent.

Stella

And my laptop Goose, my server Goose, my chat app?

Liam

Same Vilix AI memory, over MCP. One shared store, not three local ones drifting apart.

Stella

Do I run a server per machine for that?

Liam

No. Vilix AI is cloud hosted, zero infrastructure. One Vilix AI account per client keeps client work apart.

Stella

Cost to start?

Liam

Vilix AI is free forever, with a 7 day Pro trial, no credit card. Export or delete everything anytime.

Stella

My verdict: Goose remembers what you told it to remember, where you told it.

Liam

Same. Keep Goose for the work. Let Vilix AI carry the memory between sessions. Blog and guide are linked below. Does your nightly job start as a stranger? Tell us below. See you next time!