Sessions: Context and codebase understanding is not transferred between sessions on the same repository

[stevencl]

During the recent Agent Sessions UX study, participants ran into the following issue.

Problem: When a participant built up significant codebase understanding in one session (the agent reading and analysing the entire repo), they expected this understanding to carry forward to a new session on the same repository. Instead, the new session started from scratch, requiring the agent to re-analyse the entire codebase.

Observed behaviour:
P2026-03-12-1530: Spent significant time having the agent build a thorough understanding of the codebase. When they started a new session, the agent started re-analysing from scratch. Said: "I would prefer that too because there is a concept... whenever it transfers the context to another chat, it basically gives the complete knowledge, which reduces the time of the developers... It should transfer the context memory to the next step."
P2026-03-11-1400: Highlighted this as a differentiator opportunity. Said: "If you just carry forward the conversation to the next new conversation... if you are just adding that as a new feature, that will work good and comparatively better than all the AIs."

Expected behaviour:
When starting a new session on the same repository, the agent should be able to reuse prior codebase analysis.
Consider a "compact and carry forward" mechanism that transfers essential context without consuming excessive token budget.

[vs-code-engineering]

Hi @stevencl. As a member of the team, you can help us triage this issue by referring to https://github.com/microsoft/vscode-copilot/wiki/Copilot-Inbox-Tracker and assigning an owner directly.

[bamurtaugh]

Thanks for sharing this! This lack-of-context-transfer doesn't feel specific to Sessions, but I could see it feeling more prevalent when users are actively switching between sessions more often. The approach I typically use is having some sort of scratchpad markdown doc where I ask the agent to take notes on its work or findings in between sessions, and then I can cross-reference that insight in other sessions. Looping in @bhavyaus given your work on memory / planning / TODOs, and @benibenj given the Sessions angle.

[stevencl]

I think the approach you described works well, especially in a coding first experience, but I think there may well be expectations that in an agent driven experience more of this is taken care of for the user.

I think in the code focused experience, managing memory etc might be something that a code focused developer might want to do but in an agent driven experience, I think there would be value in looking at ways to at least direct users towards things like managing memory, as opposed to requiring them to know to do it, and to know how to do it.

[bhavyaus]

The only prompting Ive added to leverage memory is for Plan mode only. I have seen the agent successfully leverage session memory as a scratchpad. I haven't had the time to add additional prompts/tool use updates to instruct the agent to do this but happy to reconsider this for session based memory.