M
Memory
Stores durable decisions, assumptions, patterns, policies, lessons, and role-specific context across interactions.
05 / KNOWLEDGE
Learning over time
Remember.
Reuse.
Reconcile.
The architecture should not treat every interaction as new. Knowledge compounds when memory, cache, and governed background learning work together.
- Memory What should persist?
- Cache What can be reused?
- Dreaming What did we learn?
C
Cache
Reuses recent prompts, context, tool results, responses, searches, and retrievals to reduce latency and cost.
D
Dreaming
A governed background process that compresses, deduplicates, resolves conflicts, expires stale knowledge, and extracts patterns.
Knowledge scopes
Useful knowledge moves only through policy.
Private memories should not automatically become public. The knowledge router generalizes useful insights and verifies permissions before promotion.
Organizational memoryPolicies · enterprise decisions · common knowledge · shared lessons
Team memoryFinance · Sales · Operations · Supply Chain
Private memoryRole context · working history · specialist knowledge
Freshness is part of truth
Every cache needs a clock.
Different facts age at different speeds. Agents need source, creation time, and TTL metadata to judge whether cached knowledge is still trustworthy.
Company policy
Months
Product description
Weeks
Inventory level
Minutes
Stock price
Seconds
Incident status
Seconds
Real time + over time
Routing reacts.
Dreaming reflects.
The communication router asks, “What matters now?” Dreaming asks, “What did we learn across time?” Together they make the system responsive and cumulative.
Work
→Outcome
→Memory
→Dreaming
→Next
execution
execution
More than memory
Learning changes how agents operate.
Dreaming can update memories, skills, policies, routing rules, caches, and knowledge bases. The goal is not to remember every message. It is to make future execution better.
See learning in a scenario →