Memory Offload

clawhip × filesystem-offloaded memory

Use this skill when you want a Claw OS-style memory system where MEMORY.md stays small and points into a structured memory/ tree.

What you get

• a clear role for MEMORY.md as pointer/index/current-beliefs layer
• a practical read/write workflow for agents
• guidance for sharding memory by time, channel, project, topic, and handoff
• migration guidance away from monolithic memory files

Read order

1. Read MEMORY.md first.
2. Follow the pointer to the smallest relevant shard.
3. Read subtree indexes only when needed.
4. Avoid loading the whole memory tree by default.

Write order

1. Write detailed updates to the canonical leaf shard.
2. Update MEMORY.md only when the pointer map or current beliefs changed.
3. If a section grows noisy, split it into a dedicated file.
4. Archive cold history to keep the hot path small.

Default shard map

• memory/daily/YYYY-MM-DD.md -> chronological execution log
• memory/channels/<channel>.md -> one lane/channel
• memory/projects/<project>.md -> project/repo state
• memory/agents/<agent>.md -> agent/operator profile
• memory/topics/rules.md -> durable operating rules
• memory/topics/lessons.md -> reusable lessons
• memory/handoffs/YYYY-MM-DD-<slug>.md -> bounded handoffs
• memory/archive/YYYY-MM/ -> cold history

Offload triggers

Offload when:

• MEMORY.md stops being easy to scan
• one topic dominates the root file
• detail is only relevant to one entity or workflow
• logs or history start crowding out current beliefs

Start here

• docs/memory-offload-architecture.md
• docs/memory-offload-guide.md
• docs/examples/MEMORY.example.md
• docs/examples/memory/README.example.md