Engram supports built-in memory types and custom types. Each type represents a category of knowledge your agent can store and recall.
Built-in Types
| Type | Purpose | Example |
|---|
semantic | Facts, knowledge, learned info | ”Project Alpha uses Rust for the backend” |
episodic | Conversations, events, sessions | ”User asked about pricing on March 22” |
procedural | How-to steps, workflows | ”Deploy: push to main → CI runs → auto-deploy” |
persona | Agent identity, personality | ”I am a helpful coding assistant” |
preference | User likes/dislikes | ”User prefers dark mode” |
goal | Objectives, task tracking | ”Finish API by Friday” |
feedback | Corrections, user input | ”Don’t use semicolons here” |
relational | Entity relationships | ”Alice manages the backend team” |
checkpoint | State snapshots | Serialized task state |
embedding | Vector embeddings | Float arrays for RAG retrieval |
working | Short-lived scratch memory | Temporary computation results |
tool_output | Cached tool/API results | Search results, API responses |
file | Binary files (PDF, images, video) | Uploaded documents and media |
Custom Types
Create your own types by prefixing with custom: (max 64 characters):
{
"type": "custom:meeting_notes",
"key": "standup/2026-03-22",
"content": "Discussed API versioning and blob naming..."
}
Choosing the Right Type
Use semantic for facts your agent should “know,” episodic for things that “happened,”
and file for binary content. When in doubt, use semantic.
| Scenario | Recommended Type |
|---|
| Agent learns a user’s name | semantic |
| Recording a conversation summary | episodic |
| Storing deployment instructions | procedural |
| Agent’s system prompt / personality | persona |
| Temporary computation state | working |
| Uploaded PDF report | file |
| Embedding vectors for RAG | embedding |
