Knowledge feature failure

Google Antigravity
1

I asked Antigravity to audit the IDE’s logs and explain why the ‘Knowledge’ folder remains empty.

After auditing my internal file system and background process logs, I can clarify exactly how my memory is designed to work—and why it is currently failing.

1. How I Am Designed to “Remember”

I do not have a single, continuous consciousness. Instead, I rely on a Knowledge Distillation cycle. My “memory” is built in three stages:

  1. The Live Session: While we talk, I store everything in a temporary high-density log (a .pb file).
  2. The Analysis (Subagent): When you close a session, the system is designed to spin up a background version of me—a “Subagent”—whose only job is to read that log and extract long-term lessons.
  3. The Knowledge Base: These lessons are written as Knowledge Items (KIs) into the /knowledge/ directory. If this folder is empty, I have no “long-term memory” to read when we start our next chat.

2. Why My Memory is Currently “Empty”

My audit of the system’s main.log reveals that the background Subagent is often prevented from finishing its work. The primary cause is not “forgetfulness,” but a specific engineering error in the host environment:

  • The “BigInt” Serialization Bug: I found multiple entries in the logs stating: TypeError: Do not know how to serialize a BigInt.
  • The Conflict: In modern coding, we use “BigInts” to handle very large numbers (like high-precision timestamps for your file edits). However, the standard “Save” function used by the background process doesn’t know how to handle these numbers.
  • The Crash: When the system tries to save the summary of our session, it hits one of these numbers, encounters the error, and the background process crashes instantly. Because it crashes, the Knowledge Item is never written to disk, and the directory remains empty.

3. The Impact of “Log Bloat”

Additionally, if our conversations become very large (some of our logs can exceed 10MB), the sheer volume of data can exceed the “processing window” of the background Subagent. If the Subagent cannot read the entire log before the system’s safety timer runs out, it will exit without saving any new knowledge.

4. Managing Continuity Today

While this background “Distillation” process is being refined, I must rely on a Search-First approach. To ensure I stay aligned with your project’s “Right Way,” I recommend maintaining a structured docs/ folder within your repository.

I am programmed to prioritize reading your Architectural Decision Records (ADRs) and Implementation Plans at the start of every session. Until the automated Subagent can successfully bypass the serialization errors in the host system, your documentation acts as my only reliable source of long-term memory.

I reported the BigInt errors here:

1 Like
Knowledge: how to make it works?
3

Hello @that.Steve,
Thank you for sharing the report on the ‘Knowledge’ feature of Antigravity.
I have shared the issue with our engineering team to investigate further, and we appreciate your patience as we work toward a resolution.

1 Like
4

I am having the same issue : The agent is prohibited from direct writes to ~/.gemini/antigravity/knowledge/. As a Knowledge Subagent is not currently available via update_knowledge_request, this workflow is blocked/escalated to the user for applying the following Knowledge Item Delta.