Memory

Memory

AgentStudio keeps its own memory, shared by every backend (Claude, Codex, DeepSeek) and isolated from the machines' native CLI memory. The same memory text is injected into each backend through its natural channel, so all lanes stay on the same page across sessions.

Two tiers

• Manual memory (curated) — what you write, plus anything saved with a trigger phrase. Authoritative and always injected.
• Auto memory (learned) — silently summarized from finished conversations (Codex / Claude-Code style). Injected into later runs; editable, with Tidy and Clear.

Auto-extracted facts go into the learned tier so they never pollute your hand-written notes.

Two scopes

• Global — applies to every project. Stored under the app's userData (memory/global.md, memory/global.auto.md).
• Project — applies to the current project, stored inside the project at <project>/.agentstudio/memory.md (manual) and memory.auto.md (learned), so it travels with the repo and can be version-controlled.

Manage all four (scope × tier) in the Memory view in the sidebar.

Saving memory

Trigger phrases (manual). Start a chat message with one of these and the rest is saved straight to manual memory (no model turn runs):

记住 · 记一下 · 记下 · 别忘了 · 别忘记 · 不要忘记 · 牢记 · 务必记住 · 以后记得 · 请记住
remember · don't forget · note that · keep in mind · make a note

Example: 记住：这个项目用 pnpm，不要用 npm.

Auto extraction (learned). When a conversation finishes — a Solo turn settling, or a Dual round completing — AgentStudio asks a cheap model to extract durable facts (preferences, conventions, tech stack, decisions, pitfalls), de-duplicates against what's already stored, and appends new bullets silently. Toggle it in Settings → Memory.

Tidy & clear

In the Memory view's Auto tab:

• Tidy — runs an LLM pass to dedup/compress the learned notes (like Claude Code's "auto-dream").
• Clear — wipes that scope's auto memory.

A size cap keeps the injected memory bounded so it never blows up the context window.

Isolation from your machine

So that lanes use only AgentStudio's memory:

• Codex runs with features.memories, features.goals, and features.chronicle disabled (via -c …=false at launch only — your ~/.codex/config.toml and the 16 MB ~/.codex/memories are not modified, and your terminal Codex is unaffected).
• Claude has no machine-global memory pulled in (no ~/.claude/CLAUDE.md is injected by AgentStudio).
• DeepSeek is pure HTTP — only the injected memory reaches it.

The result is one-way isolation: AgentStudio neither reads nor writes your personal CLI memory.