Feature Request: Local auto-memory (agent-initiated, no remote storage)

[loganrosen]

Problem

Organizations with security concerns about remote data storage disable Copilot Memory (which is off by default for enterprise/org-managed subscriptions). This leaves Copilot CLI with no way to accumulate knowledge across sessions — every session starts from scratch unless the user manually maintains instruction files.

Existing feature requests (#446, #667, #1912) focus on session persistence or user-driven memory. This request is specifically about agent-initiated, local-only memory — where the agent autonomously decides what is worth remembering and writes it to local files, similar to how Claude Code auto memory works.

Proposed Feature

A local auto-memory system where Copilot CLI:

1. Autonomously writes learnings during sessions — build commands discovered, conventions learned, user corrections, debugging insights, architecture notes
2. Stores memory locally in a per-project directory (e.g., ~/.copilot/projects/<repo>/memory/), not on GitHub servers
3. Auto-loads memory into future sessions so accumulated knowledge persists
4. Uses plain markdown files that users can audit, edit, or delete at any time

How Claude Code does this (reference implementation)

Claude Code's auto memory is a good reference for what this could look like:

• Storage: ~/.claude/projects/<project>/memory/ — one directory per git repo
• Structure: A MEMORY.md index file + topic files (debugging.md, api-conventions.md, etc.)
• Loading: First 200 lines / 25KB of MEMORY.md loaded at session start; topic files read on demand
• Agent-initiated: The agent decides what to save based on whether the information would be useful in a future conversation — no explicit "remember this" needed
• Editable: Users can view, edit, or delete any memory file. Claude Code provides a /memory command to browse them.

Key design principles

• Local-only: No data leaves the machine. This addresses the security concerns that cause organizations to disable remote Copilot Memory.
• Per-project scoping: Memories are tied to a git repository, so different projects don't pollute each other.
• Agent-initiated: The agent writes notes for itself without the user having to explicitly say "remember this." When the user corrects the agent or the agent discovers something useful (e.g., "this project uses pnpm, not npm"), it saves that automatically.
• Complementary to custom instructions: CLAUDE.md / .github/copilot-instructions.md are what the user writes for the agent. Auto memory is what the agent writes for itself.
• Transparent: Plain markdown, easily auditable, with a command to browse/edit (e.g., /memory).

Why this matters

• Organizations that disable remote Copilot Memory have no alternative today. Custom instructions are manual and user-maintained — they don't capture what the agent learns during work.
• Reduces repetitive corrections: If a user corrects the agent ("use pnpm, not npm"), the agent should remember that for next time without the user having to add it to an instructions file.
• Improves over time: The agent gets better at working in a specific codebase as it accumulates project-specific knowledge.
• No security trade-off: All data stays on the local machine. Organizations don't have to choose between memory and data security.

Relationship to existing features

Feature	Who writes	Storage	Auto-initiated
.github/copilot-instructions.md	User	Repo (committed)	No
Copilot Memory (remote)	Agent	GitHub servers	Yes
This proposal (local auto-memory)	Agent	Local filesystem	Yes

This feature would fill the gap for users/orgs that want agent-initiated memory but cannot or choose not to use remote storage.