Memory

Memory is plain markdown, one MEMORY.md per scope, shown to you in the memory section of your system prompt (the first 200 lines of each file):

Scope	File	Who writes it
global	~/.pi/agent/memory/MEMORY.md	cones, for every agent
a cone (cone, cone-2, …)	~/.pi/agent/memory/<cone>/MEMORY.md	that cone
a role (role:<path>)	~/.pi/agent/agent-memory/<path>/MEMORY.md	scoops of a role with memory: { scope: user, path: <path> }

A role with memory: { scope: project, path: <path> } keeps it in <project>/.pi/agent-memory/<path>/MEMORY.md, where the project is the nearest folder above the scoop's working directory with a .pi or .git, inside /home.

Writing

Use memory_write rather than editing the files:

• section and title name the entry; saving the same section and title again replaces it.
• tag is user (who the user is), feedback (how they want you to work) or project (facts about their work).
• remove: true forgets an entry. scope: "global" writes the shared file (cones only).

Keep entries short and durable: a preference, a correction, a convention, a project fact. Don't save one-off task details, transcripts, or anything secret (keys, tokens, passwords, signed or launch URLs); memory_write redacts what looks like one. Each file holds 16 KB; when it's full, merge or remove entries before adding.

The file format, if you or the user edit it by hand:

## Preferences

### Lead with the result
tag: feedback

Keep replies short. Lead with the result, then the detail.

Extraction

Besides your own writes, SLICC mines a cone's conversation for memory when it is compacted and after it has been idle for a while (30 minutes by default), with one model call that writes into the cone's file through the same writer. "memory": { "extract": false } in ~/.pi/agent/settings.json turns it off, "extractModel": "<provider>/<model>" picks a cheaper model, and "idleMinutes" changes the wait (0 turns the idle trigger off). If the user asks why something was remembered, it may have come from extraction; edit or remove it like any entry.

The memory command

• memory scopes lists every scope with its entry count and file.
• memory show [<scope>] prints a memory file, yours by default.

The user sees and edits the same entries in the Memory panel.

Instructions are not memory

AGENTS.md and CLAUDE.md files hold instructions the user wrote: ~/.pi/agent/AGENTS.md and the ones on your working directory's path inside /home are in the project_context section. Files in mounted folders are listed there as not loaded; they load once SLICC can trust a folder.