Memory Systems
Tiered memory that lets agents keep track of the task in hand, share what they learn with each other, and recall relevant experience from your project's history.
Three Tiers
Working Memory
Immediate context for the task in hand: the conversation so far, the files being touched, the decisions already made.
- Instant access
- Scoped to the task
- Message history
- Current context
Shared Memory
State that agents share with each other and that survives restarts: maturity modes, learned patterns, coordination between agents.
- Visible to every agent
- Survives restarts and redeploys
- Pattern sharing
- Agent coordination
Semantic Memory
Knowledge retrieved by meaning rather than by exact match, so an agent finds the relevant precedent even when the wording differs.
- Meaning-based retrieval
- Similarity search
- Project knowledge base
- Scales with history
Memory Types
Short-term
Message history, current conversation context
Retention: SessionLong-term
Pattern library, learned best practices
Retention: PermanentEpisodic
Historical deployment outcomes, incident records
Retention: Fades with ageSemantic
Embedding-based knowledge for similarity search
Retention: VersionedWhat Memory Does for Your Team
Agents learn your codebase once
Conventions, hotspots and past fixes are remembered across tasks, so the tenth change to a service benefits from the first nine.
Relevant precedent is found, not searched for
An agent diagnosing an incident is shown similar past incidents and what resolved them, ranked by how relevant and how recent they are.
Stale knowledge loses weight
Older observations count for less over time and knowledge that keeps proving useful counts for more, so agents follow what works now, not what worked once.
Memory stays in your account
In a BYOC deployment all three tiers run inside your AWS account; project knowledge is scoped to your tenant and never shared across customers.