Skip to Content
DocumentationFundamentalsMental Models & Terminology

Mental Models & Terminology

Core Concepts

Memory vs. Context vs. State

Memory: Persistent information that survives across sessions

  • Stored independently of current context
  • Requires explicit retrieval
  • Can be updated, deleted, or versioned

Context: Information actively available to the model

  • Included directly in prompts
  • Immediately accessible for reasoning
  • Limited by context window size

State: Current condition or status of the agent/conversation

  • Active variables and flags
  • Session-specific data
  • Temporary working information

Retrieval vs. True Memory

Retrieval: Finding and surfacing stored information

  • Like search through a database
  • Information exists unchanged
  • Access pattern: Store → Search → Retrieve

True Memory: Dynamic, integrated knowledge system

  • Information evolves and connects
  • Context shapes retrieval
  • Access pattern: Experience → Learn → Adapt

Memory Types

Explicit Memory

Consciously stored and deliberately retrievable information.

Examples:

  • User preferences: “I prefer dark mode”
  • Historical facts: “Meeting was scheduled for 3pm”
  • Explicit instructions: “Always format dates as MM/DD/YYYY”

Implicit Memory

Learned patterns and behaviors not explicitly stored as facts.

Examples:

  • Communication style adaptation
  • Response pattern preferences
  • Interaction timing preferences

Episodic Memory

Specific events and experiences with rich contextual detail.

Structure:

  • What happened
  • When it happened
  • Who was involved
  • Where it occurred
  • Why it mattered

Semantic Memory

General knowledge and facts without specific episodic context.

Structure:

  • Facts and rules
  • Learned patterns
  • Abstracted knowledge
  • Concept relationships

Memory Operations

Encoding

Converting experiences into storable memory representations.

Challenges:

  • What information to preserve?
  • How to represent relationships?
  • Which details are significant?

Storage

Persisting memory in appropriate systems and formats.

Considerations:

  • Storage medium (SQL, vector, graph)
  • Indexing strategy
  • Compression and summarization
  • Version control

Retrieval

Finding and accessing relevant stored memories.

Types:

  • Exact match: Specific factual lookup
  • Similarity search: Semantically related content
  • Association: Connected through relationships
  • Temporal: Time-based queries

Consolidation

Strengthening, connecting, and organizing memories over time.

Processes:

  • Connecting related memories
  • Abstraction and pattern extraction
  • Importance scoring adjustment
  • Redundancy elimination

Forgetting

Deliberate removal or degradation of stored information.

Types:

  • Active deletion: Explicit removal
  • Passive decay: Automatic cleanup
  • Interference: New information overwriting old
  • Motivated forgetting: Privacy-driven removal

Entity Types

Agents

The AI system itself and other AI agents in multi-agent scenarios.

Users

Human participants with identities, preferences, and historical context.

Conversations

Interaction sessions with boundaries, topics, and outcomes.

Objects

Domain-specific entities (documents, tasks, projects, etc.).

Relationships

Connections between entities (user-to-user, user-to-object, etc.).

Temporal Models

Absolute Time

Specific dates and times with precise timestamps.

Relative Time

Time relationships without absolute anchoring (“last week”, “before the meeting”).

Event Time

Time defined by other events (“after the project started”, “since we discussed X”).

Validity Time

When information was true in the real world.

Transaction Time

When information was stored in the system.

Quality Metrics

Precision

Percentage of retrieved memories that are relevant.

Recall

Percentage of relevant memories that are retrieved.

Freshness

How current and up-to-date stored information is.

Coherence

Logical consistency between stored memories.

Coverage

Breadth of information captured about entities and relationships.

Common Anti-patterns

Memory Hoarding

Storing everything without considering relevance or importance.

Context Stuffing

Including too much historical context in prompts.

Static Snapshots

Treating memories as unchanging facts rather than evolving understanding.

Retrieval Tunnel Vision

Only using exact matches instead of exploring related information.

Temporal Flattening

Losing time-based context during storage or retrieval.

Next Steps

The Three Tiers of MemoryContext Engineering