Semantic Memory Selection
The agent doesn't remember everything -- it remembers what's relevant.
An agent can have thousands of memories stored. Injecting all of them into context on every message would be impossible and counterproductive. Semantic selection solves this problem.
When a new message arrives, the selection system analyzes its content and identifies the most relevant memories. Selection combines multiple signals: semantic proximity, recency, recall frequency, and importance.
The result is context enriched with exactly the right memories. The agent appears to remember relevant information naturally, like a human colleague who knows the project well.
Message analysis
The incoming message is transformed into a semantic vector and analyzed to extract key themes and entities.
Vector search
Semantically close memories are identified via a search in the vector database (Qdrant, pgvector).
Multi-criteria scoring
Each candidate memory receives a combined score: semantic relevance (40%), recency (20%), recall frequency (20%), importance (20%).
Selection
The top N scored memories are selected (N is configurable, typically 5-15).
Injection
Selected memories are injected into the agent's context, in a dedicated section, before the agent formulates its response.
Project continuity
A user resumes a project after a 2-week absence. On their first message mentioning the project name, the agent automatically recalls the latest technical choices.
Implicit preferences
A user mentions they're working on a React component. The agent remembers that this user prefers strict TypeScript and hooks over classes.
Multi-user
Each user has their own set of memories. When Alice mentions the dashboard, the agent remembers the EasyBoard dashboard. When Bob mentions it, it remembers the Monitoring dashboard.
featurePages.memorySelection.config
- Active by default on all EasyClaw v2 agents
- Requires a vector backend (Qdrant, pgvector, or Cognee)
- Works in synergy with Extract-Memories and Auto-Dream