Memory

Memory is a persistent, per-repository knowledge base for everything your team decides about a repo's tests — quarantine calls, assertion conventions, maintenance TODOs, and “always / never” constraints. Memories are auto-saved and auto-categorized from your Ask AI conversations (after you approve), then read back on every future turn, in the repo's Memory tab, from the CLI, and by AI coding agents over MCP.

Beta

Memory is in Beta. The Memory tab, Ask AI auto-save, and the CLI /memory command are available now; agent read/write over MCP is rolling out.

Memory vs Ask AI's conversation memory

This page is about repo Memory — a team-shared, per-repository knowledge base. It is different from Ask AI's conversation memory, which is the per-user recall of your recent chats and preferences. See Ask AI → Memory for that. Repo Memory is scoped to a repository and shared with your team, not to a person.

Why it exists

Test knowledge usually lives in one engineer's head or scrolls out of a chat history: why a test was quarantined, why an assertion is written a certain way, what still needs cleaning up. Memory captures those decisions once and keeps them durable, so neither a teammate nor an AI agent re-litigates them. Ask AI reads the repo's memory before reasoning about its tests, so its answers respect established decisions instead of contradicting them.

The four categories

Every memory is auto-classified into one of four categories so the right context surfaces at the right time.

Category	What it captures	Example
Decision	Agreed verdicts	“Quarantine login_oauth.spec — flaky in CI.”
Insight	Observations and findings	“checkout flakes on the OAuth redirect, not the payment step.”
Maintenance	Follow-up TODOs and refactors	“Split the cart suite after the v3 migration.”
Context	Constraints and conventions	“Never assert on toast timing — it's environment-dependent.”

How memories are created

There are three sources. Each memory records its source, so you always know where it came from.

From Ask AI (auto-save)

In any repo-scoped Ask AI chat, when you agree to a decision the agent proposes a memory. An Approve / Reject prompt appears above the composer — nothing is written until you confirm. On approval, the memory is saved, its category is auto-classified from the title and content, and (when relevant) it's linked to a specific test. The entry keeps a link back to the conversation that created it.

Persist on agreement

You don't need to ask the agent to “save this.” When the conversation reaches a decision, it offers to persist it — you just approve or reject.

Manually

Open the repo's Memory tab and choose Add Memory. Give it a title and content, and optionally tie it to a test. Manually created entries are marked with the Manual source.

From an AI agent over MCP (Beta)

AI coding agents can write memories over MCP (for example, recording a decision reached while fixing a test). Writes require the mcp:memory token scope. See MCP capabilities.

Viewing and managing memories

Open a repository and select the Memory tab (marked Beta). There you can:

• Scan the stats strip — total active memories, how many map to live tests, and how many reference a test spec that no longer exists.
• Spot stale entries — a memory whose linked test spec is gone is flagged with a “test spec no longer found” warning, so you can update or archive it.
• Search memories by title, content, or linked test.
• Open the conversation that created an Ask AI memory.
• Edit, archive, or delete entries. Archived memories are hidden from the agent but stay recoverable; deleting is permanent.

How memory is read back

Surface	What it does
Ask AI	Injects the repo's memory digest into every repo-scoped turn, so answers respect prior decisions.
Memory tab	The team's view — browse, search, and manage every entry.
CLI	Review and manage repo memory from the terminal with /memory. See CLI slash commands.
AI agents (MCP)	Read the digest and (Beta) write new memories over MCP. See MCP capabilities.

FAQ

FAQ

How is Memory different from Ask AI conversation memory?

Memory is a team-shared, per-repository knowledge base of test decisions, insights, maintenance notes, and constraints. Ask AI conversation memory is the per-user recall of your recent chats and preferences. Memory is scoped to a repository and shared with the whole team; conversation memory is scoped to you.

Will the agent save things without asking?

No. In the web app, a write pauses for an Approve / Reject confirmation above the composer before anything is saved. You stay in control of what becomes durable memory.

What happens to a memory if its linked test is deleted?

The memory is kept but flagged with a stale-test-spec warning in the Memory tab and in the digest, so you can update the link or archive it.

Who can see a repository’s memory?

Everyone with access to that repository in your organization. Memory is team-shared by design — that is what stops the same decisions from being re-litigated.

Related

• Ask AI — where most memories are created (and the home of conversation memory).
• CLI slash commands — manage memory from the terminal with /memory.
• MCP capabilities — how AI agents read and write memory.
• Repositories — where the Memory tab lives.