What the Fluctuating Memory Metric in Gemini CLI Really Means

When Gemini CLI runs with experimental features, the status bar shows a memory: XXX MB metric that can range from a few hundred megabytes to over 200 MB and fluctuates during a conversation.

Core definition: physical size of the working memory

The displayed Memory value is the serialized physical size of the current Active Context , i.e., the data the CLI prepares to send to the Gemini model.

Active Context consists of:

Multi‑level directives : GEMINI.md files from global, project, and subdirectory scopes.

Tool output : thousands of lines of grep results, file contents, and shell command output.

Multimodal data : uploaded screenshots or images.

Skills & MCP data : 33 loaded skills and their associated API documentation.

When these pieces are bundled for the model, the CLI computes their total physical volume; a value such as 152.8 MB indicates a substantial amount of background information is currently loaded.

Why the number shrinks: on‑demand loading and automatic compression

Two techniques drive the observed decrease after work begins.

1. jitContext (Just‑In‑Time Context)

Loading all project files would exceed token limits. Enabling jitContext makes the CLI load only the GEMINI.md files that are immediately relevant. When the user switches to a different subdirectory or changes task goals, the CLI dynamically unloads irrelevant background directives and loads the needed ones, causing the Memory metric to rise and fall.

2. Context Compaction

To save token costs and improve response speed, Gemini CLI includes a “road‑roller” mechanism that automatically detects redundant information in the conversation history. It either summarizes large intermediate log outputs or prunes them outright, cleaning out outdated process data while the discussion continues.

Memory V2: managing memory like code

The configuration flag memoryV2: true activates Gemini CLI’s latest memory architecture, replacing an opaque database with a transparent, Git‑trackable Markdown file system.

The four‑layer cascade consists of:

Global : global preferences (e.g., “I prefer TypeScript”).

Project : project‑specific conventions (e.g., “this repo uses Vitest”).

Subdirectory : knowledge scoped to sub‑modules.

Private : local private notes (e.g., local test environment IP).

Common management commands

Interactive commands to inspect or adjust the current context: /memory show: displays all directives and facts visible to the model. /memory inbox: shows the autoMemory snippets automatically extracted as experience fragments. /memory prune: manually clears persisted memory that is no longer needed.

Conclusion

The fluctuating MB value reflects Gemini CLI’s dynamic management of context health. It is not a traditional resource‑usage metric but an indicator of information throughput between the user and the AI. When the number drops, the CLI has successfully trimmed unnecessary knowledge, allowing the model to focus on the current problem.