Skills & Memory

Skills and memory are two systems that give your agents specialized capabilities and persistent knowledge across conversations.

Skills

Skills are pre-packaged sets of instructions that teach your agent specific workflows and capabilities. They are Markdown files deployed to the agent’s workspace.

How skills work

Each skill is a Markdown file (SKILL.md) stored in the agent’s workspace at workspace/skills/<skill-name>/SKILL.md. The agent runtime:

1. Scans the skills/ directory for skill directories containing a SKILL.md file.
2. Extracts a name and one-line description from each skill file (supports both YAML frontmatter and legacy heading formats).
3. Injects a summary list of all available skills into the agent’s context.
4. When a skill is relevant, the agent reads the full SKILL.md file for detailed instructions.

This keeps the context window efficient while giving the agent access to detailed knowledge on demand.

Selecting skills

Skills are chosen during agent creation (Step 3) or in the edit screen (Skills tab). They are organized into categories:

• Developer — Coding, debugging, and software engineering workflows
• Productivity — Task management, writing, and organizational skills
• Web3 — Blockchain, smart contract, and decentralized application skills

Skills are optional — if you don’t select any, your agent uses its default capabilities (bash terminal, file system access, web tools).

Skills gallery

The Gallery tab on the main dashboard lets you browse all available templates and skills. Each template card shows:

• Template name and description
• Category (with color coding)
• Model used
• Number of skills included

Tapping a template takes you to the agent creation wizard with that template pre-selected.

Memory

Memory gives your agent persistent knowledge that survives across conversations. The agent reads and writes memory files in its workspace.

How memory works

The agent’s memory system consists of:

• workspace/memory/MEMORY.md — The agent’s long-term memory file. Its contents are always loaded into the agent’s context at the start of each conversation. The agent organizes this file by topic rather than chronologically.
• Daily notes (workspace/memory/YYYY-MM-DD.md) — Dated files for session-specific context. Only today’s daily notes file is loaded into the context, keeping older notes accessible but out of the way.

The agent reads its memory at the start of each conversation and can update it by writing to the memory files using its file tools. This allows the agent to:

• Remember user preferences
• Track ongoing projects
• Store solutions to recurring problems
• Maintain notes across sessions

Context building

When a conversation starts, the agent’s context is assembled in two parts to optimize inference performance:

Static portion (cacheable):

1. Agent identity (name, date, workspace path, available tools)
2. Your system prompt (the instructions you wrote)
3. Memory system instructions

Dynamic portion (per-turn):

1. Long-term memory (MEMORY.md contents)
2. Today’s daily notes
3. Skill summaries (name + description for each assigned skill)

The static portion uses a date-only timestamp (not a full timestamp) so it remains identical across turns within a day. This allows the inference engine to cache and reuse computations for the static prefix, making responses faster.

Context compaction

When conversations grow long, the agent uses context compaction to stay within the model’s token limit. Old messages are summarized by the LLM while recent messages are kept intact. This preserves the most relevant context while making room for new interactions.