GitHub has introduced a three‑part intelligence layer for Copilot—Memory, Dreaming, and Code Search—now generally available or in public preview. The change adds persistent context, automated memory curation, and organization‑wide code discovery, directly affecting how AI‑assisted tools integrate into development pipelines.
Memory: Persistent Preferences Across Sessions
Copilot Memory is GA for paid Copilot plans. It records personal preferences (e.g., start pull‑request descriptions with a TL;DR) and repository‑level conventions (such as test commands or review expectations) and reuses them in any Copilot client, including the Copilot app, CLI, VS Code, cloud agent, code‑review UI, Slack, Teams, and Mobile. Users can create, view, or delete memories via the Copilot settings UI.
During the public preview, 93 % of reviewed CLI sessions that used memory showed better alignment with developer preferences, and 42 % demonstrated more efficient workflows attributed to the added context.
Dreaming: Automated Context Refinement
Dreaming is a preview feature that reflects on past interactions to update memories automatically. Three dream types are offered:
- Personal memory curation – adjusts individual preferences based on conversation history.
- Repository customization – proposes updates to repository‑scoped memories, skills, and custom instruction files.
- Code review customization – converts recurring accepted review feedback into guidance for Copilot code review.
Dreams generate pull‑request proposals for team review. The feature is opt‑in; individual users enable personal dreaming in their settings, while organizations activate it through Copilot policies. Dream runs consume AI credits, and administrators can delegate policy control to organization admins.
Code Search: Cross‑Repository Context
GitHub code search now feeds Copilot with the ability to locate code, configuration, schemas, and documentation across any repository in an organization without a local clone. Searches support exact text, regular expressions, and semantic queries that describe behavior (e.g., “retry failed payment”). Users invoke the feature with /research prompts such as “How do these components retry failed charges, and where could retries overlap?”
Practical scenarios include:
- Planning a migration by identifying implementations and callers across services and flagging remaining API usage.
- Investigating a bug by tracing related implementations across shared libraries and services.
- Ensuring compliance with organizational standards by surfacing existing patterns before modifying unfamiliar code.
This reduces the need to clone multiple repos, manually locate files, or copy code snippets into a conversation, while giving Copilot richer context for generating accurate changes.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Engineers should evaluate enabling Memory on paid Copilot subscriptions to capture team conventions and reduce repetitive prompting. Security and compliance teams need to audit the scope of repository memories and ensure that access controls align with existing policies, since memories can be created automatically.
Organizations considering Dreaming must decide whether to adopt the opt‑in model and allocate AI credits for automated memory updates. Administrators should configure Copilot policies to govern who can enable Dreaming and to which repositories.
Integrating Code Search into daily workflows can streamline cross‑repo investigations and migration planning. Teams may want to update internal documentation to reference the /research command and to define naming conventions for semantic queries.
Overall, the new intelligence layer encourages a shift from ad‑hoc prompting toward a more stateful, context‑aware AI assistant. Practitioners should pilot the features in a low‑risk repository, monitor memory usage and credit consumption, and establish governance around memory lifecycle and Dreaming proposals before scaling organization‑wide.
