SIGNAL / TOOLS

GitHub Copilot Memory and Dreaming: what's GA, what's preview

SOURCEGitHub

PUBLISHEDOctober 28, 2026

LAYERTools

SIGNAL SCORE28/100

SXF VERIFIED BRIEF

GitHub's October 28 intelligence launch combines three related but distinct capabilities: Copilot Memory is generally available on paid plans, Copilot Dreaming is a public preview, and contextual code search helps agents find relevant code and history across repository boundaries. Memory can reuse personal preferences and repository conventions between sessions rather than starting from scratch each time.

WHY IT MATTERS

Developers repeatedly spend time giving coding agents the same test commands, architecture assumptions and team conventions. Shared, reviewable repository memories and search across codebases may reduce that repeated setup. Dreaming also raises a governance question: automatically refreshed instructions can become stale or incorrect, so review boundaries and access permissions matter as much as convenience.

WHAT TO VERIFY

Confirm the paid plan and enterprise policies for your Copilot surfaces; memory availability and settings can differ across clients. Dreaming requires opt-in and is still a preview. Review which repositories Copilot can access, whether instruction changes arrive as pull requests, and whether any vendor-reported improvement percentages generalize to your team's workflow.

SXF RESEARCH EXPLAINER

What the technical evidence means.

Source-grounded analysis

TECHNICAL ANALYSIS / 01

Memory versus Dreaming versus code search

Copilot Memory stores what an agent has learned about a project or developer. Repository memories concern conventions such as which command runs tests or where a service's configuration lives; personal memories capture a developer's own preferences. The central promise is continuity between sessions and, subject to policy and surface availability, across Copilot products. That is different from the model permanently learning new weights.

Dreaming is a separate process for maintaining context. GitHub says it can examine eligible previous sessions and accepted review feedback, update stale personal or repository memories, and propose changes to checked-in skills or instructions through pull requests. Code search addresses another problem: locating symbols, documentation and implementation patterns across services without cloning every repository. A team may need one capability without enabling the others.

Check primary-source evidence ↗

TECHNICAL ANALYSIS / 02

What changed for coding teams on October 28

The GA milestone matters because paid Copilot users can evaluate memory as a normal workflow capability rather than as a limited experiment. A developer can ask Copilot to respect repository tests or coding conventions, then inspect whether subsequent sessions follow those details. Organization and enterprise policy remains a separate layer: paying for a plan does not guarantee that every workspace or surface permits all memory types.

The release also makes claims about preview outcomes. GitHub says 93% of the reviewed Copilot CLI sessions that used memory showed better adherence to preferences and 42% showed more efficient workflows. Because the figures are drawn from a vendor-reviewed subset of preview sessions, they are useful hypotheses for an internal trial, not proof of universal speedups. A responsible adoption test should record the baseline and a repeatable task set.

Check primary-source evidence ↗

TECHNICAL ANALYSIS / 03

Privacy, policy and repository safety

Treat memories as editable operational context rather than an authority on repository truth. They may reflect outdated test scripts, misunderstood code paths or preferences that should not be shared with a team. GitHub distinguishes personal preferences from repository-wide facts, and administrator access controls matter where private repositories or organizational policy are involved. Review current memory and dreaming settings for the specific subscription and client.

Copilot Dreaming is opt-in and still in public preview. If it proposes updates to checked-in skills or instructions, those changes should receive the same human pull-request review as any other change to developer workflow controls. Measure false memory reuse, outdated instructions, privacy boundaries and tool-access scope alongside coding completion metrics before enabling it across a company.

Check primary-source evidence ↗

PRIMARY-SOURCE FACTS

What the evidence establishes.

Reviewed 2026-11-02

GA versus public preview

GitHub's October 28 release states Copilot Memory is generally available on paid Copilot plans while Copilot Dreaming is in public preview. Primary evidence ↗

Two kinds of memory

GitHub documentation describes repository-level facts and user-level personal preferences, with distinct scope and controls. Primary evidence ↗

Dreaming workflow

Dreaming can curate memories and propose changes to checked-in repository skills or instructions for review. Primary evidence ↗

Code search

GitHub describes exact, regex and semantic cross-repository search to give agents code and development history relevant to a task. Primary evidence ↗

Vendor preview observations

GitHub reports 93% improved preference adherence and 42% more efficient workflows among reviewed CLI preview sessions using memory; these are vendor-observed figures, not an independent benchmark. Primary evidence ↗

LIMITATIONS

The GA statement applies to Copilot Memory on eligible paid plans, not Copilot Dreaming, which remains a public preview. Feature access varies by settings, product surface and enterprise policies. GitHub's reported 93% and 42% observations are not controlled independent benchmarks and should not be presented as guaranteed gains.

TOPICSGitHub Copilot Coding AI

MODELSNo named model detected

GitHub Copilot Memory and Dreaming: what's GA, what's preview

What the technical evidence means.

Memory versus Dreaming versus code search

What changed for coding teams on October 28

Privacy, policy and repository safety

What the evidence establishes.

Keep the context connected.

Route and orchestrate models with GitHub Copilot Fusion

GPT-6.1 Terra in GitHub Copilot

GitHub Copilot models deprecated on October 19, 2026

Upcoming deprecation of selected GitHub Copilot models in mid-November

Gemini 4 Argon in GitHub Copilot