Concepts

Concepts are curated knowledge — generalized understanding that the agent creates from episodes. They're the persistent, meaningful output of Remind's memory system.

Structure

Each concept has:

Field	Description
id	Unique identifier
title	Short descriptive title
summary	Natural language description of the knowledge
concept_type	Type: pattern (generalizations), fact_cluster (grouped facts)
confidence	How certain (0.0–1.0)
instance_count	How many episodes support this concept
relations	Typed edges to other concepts
conditions	When/where this concept applies
exceptions	Known cases where it doesn't hold
source_episodes	Episode IDs this was derived from
specifics	For fact_clusters: list of verbatim facts
entity_ids	Entities linked to this concept
embedding	Dense vector for similarity retrieval
tags	Searchable tags
topic	Knowledge area (e.g., "architecture", "product")
decay_factor	Retrieval priority weight (affected by memory decay)

How concepts differ from episodes

	Episodes	Concepts
Nature	Specific, raw	Curated, structured
Created by	remember()	Agent via apply
Lifespan	Pending → processed	Persistent
Structure	Plain text + type	Text + confidence + relations
Retrieval	Direct search	Primary retrieval target

Concept types

Pattern concepts

Generalizations the agent creates from observations, decisions, and outcomes:

Pattern: "The team favors statically typed languages for backend services"

• Created from multiple language preference observations
• Generalizes across specific instances

concept from=ep:11,ep:12,ep:13 title="Language preferences" "Team favors statically typed languages"

Fact clusters

Groups of related concrete facts, created automatically when fact episodes are stored:

Fact Cluster: "Redis Configuration"

• Redis TTL: 300s (alice, 2024-06-01)
• Redis connection pool: 10 (bob, 2024-06-05)

Fact clusters preserve exact values. When facts conflict, both are kept and a conflict is opened.

Standalone facts

Single fact episodes that don't share entities with other facts are not clustered. They remain as first-class retrieval targets via direct episode search.

Creating concepts with apply

Agents create concepts via the apply tool:

concept as=c1 from=ep:11,ep:12 title="Auth pattern" "JWT with refresh tokens"
link source=$c1 type=implies target=concept:security
processed ids=ep:11,ep:12

This:

1. Creates a concept from episodes 11 and 12
2. Links it to an existing security concept
3. Marks the source episodes as processed

Concept quality

Good concepts are:

• Specific and falsifiable — Making concrete claims, not abstract platitudes
• Generalized (for patterns) — Not just restating a single episode
• Verbatim (for facts) — Preserving exact values
• Conditional — Stating when they apply
• Connected — Linked to related concepts via typed relations
• Grounded — Traceable back to source episodes

Topics

Concepts can belong to a topic — a knowledge area like "architecture" or "product". Topics scope retrieval (queries can be filtered to a specific topic).

Concepts without a topic are treated as general knowledge accessible from any topic context.

Managing concepts

# View a concept
remind snapshot concept:<id>

# Create a concept (via apply)
remind apply << 'EOF'
concept from=ep:11,ep:12 title="New pattern" "Description of the pattern"
EOF

# Update a concept's summary
remind update-concept <id> -s "Refined understanding"

# Update confidence
remind update-concept <id> --confidence 0.9

# Reassign or clear topic
remind update-concept <id> --topic product
remind update-concept <id> --clear-topic

# Soft delete
remind delete-concept <id>

# Restore
remind restore-concept <id>

TIP

Updating a concept's summary clears its embedding vector. The embedding is regenerated on the next recall query.