Does Roo-Code support long-term memory (like mem0) across sessions?
Can Roo-Code learn from past interactions through a long-term memory provider such as mem0? I filed a feature request in May 2025 proposing that Roo save the final completion message as a memory via the mem0 API and retrieve it on new tasks, but the issue was closed as NOT_PLANNED on 2025-09-16. Is there any official support or workaround for cross-session memory in Roo-Code?
Originally asked on GitHub: https://github.com/RooCodeInc/Roo-Code/issues/3489
4 answers
Status: Not planned. Roo-Code closed this feature request on 2025-09-16 (marked NOT_PLANNED), and a contributor stated in June 2025 that memory-based solutions are not being pursued at this time.
The reason given: long-term memory systems risk accumulating outdated or incorrect assumptions, which can degrade accuracy over time. The project's stated priority is codebase indexing, which keeps context fresh and grounded in the actual source code. As of July 2025 there was no pull request for this feature; a contributor said it is "on our radar" but gave no timeline.
For a workaround, one contributor reported experimenting with mem0 and finding that it either stored too little or too much, polluting the context. Their conclusion was that instructing the model on how a task should be done beats digging through failed scenarios. Another contributor suggested an alternative design — storing memories as plain files (title + content) in a hidden folder and reusing the code-indexing pipeline — but that was a proposal, not an implementation.
In short: cross-session long-term memory via mem0 (or any provider) is not supported today, and the team has explicitly deprioritized it in favor of codebase indexing.
Did this work for you?What @kvokka tried on GitHub (Didn't work)
A contributor tested mem0 for storing task memories and reported that it either stored too little or too much, polluting the context. Their conclusion was that it is better to tell the model how something should be done than to search through failed scenarios.
Source: https://github.com/RooCodeInc/Roo-Code/issues/3489#issuecomment-2888498519 (opens external site)
Did this work for you?What @dlab-anton tried on GitHub (Outcome unclear)
A contributor ran a real task and prompted attempt_completion using the issue's memory constraints as an example. The comment shows three example memory entries (e.g., a lesson about keeping Alembic migrations in sync with SQLAlchemy models) and notes four memories would have been created from the task. This was a demonstration of the proposed prompt format, not a validated implementation.
Source: https://github.com/RooCodeInc/Roo-Code/issues/3489#issuecomment-2878846996 (opens external site)
Did this work for you?What @tombedor tried on GitHub (Outcome unclear)
A commenter who built a CLI assistant with memory proposed storing memories as files (title + content) in a hidden folder and reusing the code-indexing pipeline, with recall tied to task lifecycle events. This was a design proposal, not an implementation.
Source: https://github.com/RooCodeInc/Roo-Code/issues/3489#issuecomment-2940631905 (opens external site)
Did this work for you?