GlossaryThe vocabulary of
The vocabulary of
AI memory.
Plain definitions for the concepts behind governed, persistent memory for AI agents.
- Agent memory
- The stored context an AI agent can save and retrieve across a workflow. Agent memory can include durable knowledge, such as decisions and constraints, and temporary working state that expires.
- Persistent memory
- Memory that remains available after a session ends. Persistent memory gives connected agents a place to keep useful knowledge for later retrieval rather than losing it when the session closes.
- Working memory (session state)
- Temporary context scoped to a session with a configurable expiry. Working memory holds in-progress state, while persistent memory holds knowledge intended to outlive the session.
- Claim lifecycle
- The set of states a stored claim moves through as knowledge changes: candidate, confirmed, stale, superseded, conflicted, and archived. The lifecycle keeps changing knowledge under review instead of treating every stored item as equally current.
- Provenance
- The source context attached to a stored memory. Provenance records where a claim came from so it can be traced, reviewed, and trusted appropriately.
- Memory governance
- The policies that control what enters memory, how conflicts are handled, and how recall is scoped. Governance is applied by a control plane separate from model reasoning.
- Retrieval-augmented generation (RAG)
- A pattern that retrieves selected information and places it into a model's context for a single response. RAG supplies context for a request; persistent memory retains information so it remains available across sessions.
- Semantic search
- Retrieval based on meaning rather than exact wording. Semantic search helps surface related information when the words in a question differ from the words in a stored memory.
- Hybrid search
- Retrieval that combines semantic matching with keyword matching. Hybrid search supports both meaning-based recall and precise terms.
- Embedding
- A numeric representation of content used for semantic comparison. Embeddings let a memory system measure similarity between a query and stored knowledge.
- Model Context Protocol (MCP)
- An open protocol that lets compatible AI clients call tools. Recall Racoon exposes memory operations through MCP so connected agents can save and retrieve memory within a workflow.
- Namespace
- A scope that separates stored memory by user, project, or session. Namespaces keep retrieval within the intended context rather than a shared pool.
- Conflict review
- The process of surfacing conflicting information and routing it for resolution, including human review when needed, rather than silently choosing one version.
- Credential vault
- An encrypted store for detected secrets. When the vault is available, detected values are stored encrypted and replaced with references in memory; if it is unavailable, detected values are redacted.
- Memory poisoning
- An attack or failure mode in which incorrect or malicious content enters memory and later influences an agent's behavior. Governance, provenance, and conflict review reduce the risk of poisoned memory being trusted.
Go DeeperConcepts into
Concepts into
practice.
See how these concepts come together in the product, or compare persistent memory with retrieval-augmented generation.