Does the Gemini API remember previous conversations?

Short answer

Not on its own. With the legacy generateContent API, every call is stateless and your code resends the conversation history. The Interactions API can keep that history server-side through previous_interaction_id, stored 55 days on the paid tier and 1 day on the free tier. Neither remembers facts across conversations; for that you build memory yourself or use Memory Bank on Google Cloud's Agent Platform.

Three things people call “memory”

When developers ask whether the Gemini API has memory, they usually mean one of three things. Only the first is built in. As of November 2026:

What you want What Gemini offers Built in?
The model remembers earlier turns in the same conversation Interactions API with previous_interaction_id, or resending history with generateContent Yes, per conversation
Cheaper, faster repeated context Implicit and explicit context caching Yes, but it is not memory
The agent remembers a user or a team across conversations Memory Bank on Google Cloud’s Agent Platform, or your own store No, you build or add it

Conversation state: generateContent versus Interactions

The original generateContent API, which Google now calls legacy but still fully supports, is stateless. A multi-turn chat works by sending the full conversation history with every request. The SDK’s chat helpers do that for you, but the history lives in your application, not at Google.

The Interactions API, generally available since June 2026 and now Google’s recommended interface, can keep state on the server:

  1. Make a call. By default the request is stored (store=true).
  2. For the next turn, pass the ID of the completed interaction as previous_interaction_id. The server retrieves the history; you do not resend it.
  3. Send tools, system_instruction and generation_config again each time. Only the conversation history carries over.

How long Google keeps stored interactions:

Tier Retention
Paid 55 days by default; you can set 7, 14, 28 or 55 days in AI Studio
Free 1 day

Set store=false to opt out. Then previous_interaction_id and background execution no longer work, and you are back to sending history yourself.

The common mistake: caching is not memory

Context caching sounds like memory, but it is a cost optimization. Google describes it as a way to reuse repeated input tokens. Implicit caching is on by default for Gemini 2.5 and newer models, with a minimum of 4,096 input tokens on Gemini 4 models and 2,048 on Gemini 2.5. When a request hits the cache, the savings are passed on, and you can see cached tokens in usage.total_cached_tokens. Explicit caching, where you create cache objects yourself, is not supported in the Interactions API or for Gemini 4 models.

Neither changes what the model knows. If the information is not in the request or the stored conversation, Gemini does not have it.

Memory across conversations

For an agent that remembers a user next week, Google offers Memory Bank on Gemini Enterprise Agent Platform, the evolution of Vertex AI. It stores long-term user preferences and facts generated from sessions, uses generative models to extract those memories, and lets an agent fetch them later. Otherwise you build it yourself: a database or vector store your code reads before each call and writes to afterwards.

If data retention matters, check the zero data retention settings on Google Cloud: caching, request logging, store in the Interactions API and some grounding features each keep data in their own way, and some cannot be switched off.

What this means for a team

Every call starts from what you send. In a team that builds several agents, or uses Gemini next to Claude Code, Cursor or Copilot, that means every agent has its own history and its own store, if it has one at all. A fact that one agent learned is not available to the next unless someone writes it somewhere they both read. An interaction stored for 55 days does not turn into lasting knowledge either. Related: what is an MCP memory server and types of AI agent memory.

Where Bombos fits

If you are about to build a memory store for your Gemini agents, Bombos can be that layer: a plugin that gives every AI tool and agent your team uses one shared memory, with team knowledge, rules learned from corrections, and what already sits in mail, documents and repos. It is a context graph with retrieval, not a model, so you can see what each answer rests on, in the cloud or on your own server. One developer can try it first with their own agents.

Related questions

Is context caching a form of memory?

No. Caching reuses repeated input tokens to cut cost and latency. Implicit caching is on by default for Gemini 2.5 and newer models, but the model only sees what you send in the request.

Do tools and system instructions carry over with previous_interaction_id?

No. Only the conversation history carries over. Tools, system_instruction and generation_config are scoped to one interaction, so you send them again each turn.

Can I stop Google from storing my interactions?

Yes, set store=false. That also means you cannot use previous_interaction_id for the next turn, and background execution does not work.

Does the Gemini app's memory work through the API?

No. Memory in the Gemini app belongs to a personal Google account. The API starts from nothing apart from what your code sends or what you have stored in an interaction.

Sources

  1. Google AI Studio docs: Interactions API overview
  2. Google AI Studio docs: Context caching
  3. Google Cloud: Memory Bank overview (Agent Platform)
  4. Google Cloud: Agent Platform overview
  5. Google Cloud: data governance and zero data retention
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.