Memory Types

Engram automatically classifies every memory into one of five categories. Classification happens at ingestion time through the /v1/intelligence and /v1/store endpoints, so you never need to manually sort memories.

Categories

preference

User preferences, opinions, and choices. These shape how an AI assistant behaves.

• "I prefer dark mode"
• "Use TypeScript over JavaScript"
• "Always respond in bullet points"

decision

Decisions made by a person or team. Captures the what and often the why.

• "We chose the relational database for this project"
• "Migrating to cloud next quarter"
• "Deprecating the v1 endpoint in June"

fact

Objective, verifiable information. Reference data that doesn't change often.

• "The API rate limit is 100 req/min"
• "Python 3.12 released Oct 2024"
• "The office is at 123 Main St"

entity

People, systems, places, and their relationships.

• "John is the CTO at Acme Corp"
• "Brain 3 is at 10.0.0.12"
• "The staging cluster runs on k3s"

other

Everything that doesn't fit the above categories. Engram still stores, embeds, and indexes these memories identically.

Automatic Classification

When you store a memory, Engram classifies it automatically:

curl -X POST https://api.engrammemory.ai/v1/store \
  -H "Authorization: Bearer pr_live_xxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "We decided to use a caching layer for session management"
  }'

Response:

{
  "id": "mem_8f3a1b2c",
  "text": "We decided to use a caching layer for session management",
  "category": "decision",
  "importance": 0.7,
  "created_at": "2026-04-10T14:30:00Z"
}

Overriding Classification

Pass the category parameter to override auto-classification:

curl -X POST https://api.engrammemory.ai/v1/store \
  -H "Authorization: Bearer pr_live_xxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "The caching layer stores data in memory for fast access",
    "category": "fact"
  }'

from engrammemory import Engram

client = Engram(api_key="pr_live_xxxxx")

# Auto-classified
memory = client.store("We chose the relational database for this project")
print(memory.category)  # "decision"

# Manual override
memory = client.store(
    "The caching layer stores data in memory for fast access",
    category="fact"
)

import { Engram } from 'engrammemory-ai';

const engram = new Engram({ apiKey: 'pr_live_xxxxx' });

// Auto-classified
const memory = await engram.store('We chose the relational database for this project');
console.log(memory.category); // "decision"

// Manual override
const memory2 = await engram.store('The caching layer stores data in memory for fast access', {
  category: 'fact',
});

Importance Scoring

Every memory receives an importance score between 0.0 and 1.0. This float value affects two things:

1. Recall priority -- Higher-importance memories rank higher in search results when relevance scores are close.
2. Lifecycle decay -- Lower-importance memories decay faster and may be archived or pruned sooner.

You can set importance explicitly:

curl -X POST https://api.engrammemory.ai/v1/store \
  -H "Authorization: Bearer pr_live_xxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Production database credentials rotated on April 1",
    "importance": 0.95
  }'

When omitted, Engram assigns importance automatically based on the content's specificity, actionability, and category.