Memory
Clawpy uses a multi-layer memory architecture that gives agents both short-term working context and long-term persistent knowledge. This is what separates Clawpy from stateless chat assistants — your agents genuinely remember and learn.
Two Memory Systems
Clawpy runs two distinct memory systems side by side:
Alfred (Personal Memory)
Alfred is your virtual butler agent. He maintains a personal relationship with you across sessions:
- Conversations — Your full chat history, retained for up to 3 years with an access-aware decay model. Conversations you revisit frequently last longer.
- Learned facts — Your name, preferences, project details, and corrections you have made. These are permanently stored and automatically injected into every conversation.
Alfred learns passively. Every few exchanges, a background process extracts durable facts (your name, your projects, your preferences) and pins them in memory. You never need to explicitly "teach" him.
Workspace Memory (Task Memory)
While Alfred remembers you, workspace agents remember the work. Their memory is organised in layers:
| Layer | What It Stores | Lifespan |
|---|---|---|
| Session | The current conversation context | Until the session ends |
| Daily Notes | A summary of what happened today | Configurable |
| Durable Knowledge | Project facts, architecture decisions, canonical truths | Configurable |
| Vector Search | Semantic embeddings for cross-session recall | Configurable |
Retention Tiers
You can configure how long workspace agents remember things:
- Goldfish — 8 hours (useful for throwaway tasks)
- Standard — 7 days
- Elephant — 90 days
- Permanent — Never forget
Set the retention tier in each agent's configuration file.
How Recall Works
When you ask an agent a question, it searches both its factual knowledge and past conversations for relevant context. Matched results are injected into the system prompt, giving the LLM full awareness of prior work.
This means an agent can recall "that architecture decision we made last week" or "the bug we fixed in the API layer yesterday" without you needing to repeat context.
Managing Memory
- Clear conversation history — Use the trash button in the Alfred interface. This clears chat logs but preserves learned facts.
- Correct facts — Simply tell Alfred the correct information. He will extract the correction and update his knowledge.
- Full reset — Available through the settings panel for a complete memory wipe.