Memories
Memories store context learned from previous runs. They let Agents reuse what worked, avoid repeated mistakes, and adapt to your team's preferences over time.
Memories include specific events, generalized patterns, and reusable strategies gathered during real workflows. Memories that are no longer accessed lose relevance over time.
Memories versus knowledge base
Memories are not the same thing as the knowledge base.
| Source | What it contains | How it gets there |
|---|---|---|
| Knowledge base | Notes, files, and synced repositories you save for assistants | Your team adds entries or connects GitHub repositories |
| Memories | Learned observations, facts, and working patterns | Agents create and refine them while doing work |
| Organization context | Broad background such as products, audience, and internal terminology | Your team writes and maintains the shared description |
Use the knowledge base when you want to give assistants durable reference material. You can add individual files or connect a GitHub repository so Altertable syncs Markdown, text files, and dbt manifest metadata from a branch. Let memories capture what the system learns from actual runs, feedback, and repeated patterns.
Goals of Memories
- Remember your preferences: metric definitions, preferred insight types, important KPIs
- Learn from feedback: when you upvote or downvote Notifications, agents remember what matters
- Build business context: industry-specific terminology, seasonal patterns, key stakeholders
- Avoid repeating mistakes: lessons from past analyses carry forward
- Retain entity knowledge: facts about specific customers, products, or segments persist across runs
Memory Types
Agents structure their knowledge into three categories, mirroring how people naturally learn:
| Type | In short | Description | Example |
|---|---|---|---|
| Episodic | "What happened" | Specific events and experiences tied to a moment | "Revenue dropped 15% during the July migration" |
| Semantic | "What I know" | General facts and patterns derived from experience | "Revenue typically dips in Q3 due to seasonal trends" |
| Procedural | "How to do it" | Techniques and best practices learned through practice | "When analyzing churn, always segment by acquisition channel" |
You experience something (episodic), extract a general lesson from it (semantic), and develop a reliable approach for next time (procedural).
Memory Sources
| Source | Description |
|---|---|
| Agent runs | Agents create memories as they analyze data, detect patterns, and generate notifications |
| Notification reviews | When you upvote or downvote a Notification, the agent remembers your decision |
Memory Lifecycle
Memories go through four stages:
| Stage | Description |
|---|---|
| Created | An agent records an observation during its work |
| Active | The memory is retrieved and used by agents during their work |
| Consolidated | Similar episodic memories are grouped and distilled into semantic or procedural knowledge |
| Faded | Memories that lose interest gradually decay and are eventually removed |
For example, an agent might record several episodic memories about revenue drops during different months. Over time, these are consolidated into a single semantic memory: "Revenue typically dips during infrastructure migrations." The original details fade while the pattern persists.
Memory Scopes
Memories are organized into a hierarchy of scopes, from organization-wide knowledge shared across all agents down to personal preferences visible only to a specific team member. This lets agents apply the right context at the right level, whether it's a company-wide fact or a detail about a specific business entity.
When agents detect the same pattern across multiple workflows, they can promote it so all agents benefit. At the other end, personal memories keep each team member's preferences private and separate.
| Scope | What it is for |
|---|---|
| Organization | Company-wide learnings |
| Workflow | Task-specific knowledge |
| Agent | Role-specific patterns |
| Entity | Entity-specific context |
| User | User-specific context |
Memories are managed per environment. In organization settings, open the memories list for an environment to search, edit, or remove saved memories.
Relevance and Retention
Agents don't keep everything forever. Each memory has a relevance score driven by importance, recency, and frequency of access. Memories that are frequently used retain their relevance. Memories that stop being accessed gradually decay and are eventually removed.
A configurable decay rate controls how quickly unused memories fade:
| Rate | Best for |
|---|---|
| Daily | Transient context that becomes stale quickly |
| Weekly | Standard observations and patterns (default) |
| Monthly | Long-lived business knowledge and user preferences |
Learn More
- Agents: the workers that create and use memories
- Notifications: findings that agents generate using their accumulated knowledge
- Insights: visualizations and analyses that agents remember context about
- Dashboards: monitor metrics with agents that remember your context