Typed Memory

Aegis Memory v1.9.0 introduces 4 cognitive memory types inspired by research SOTA systems (MIRIX, G-Memory, BMAM). These types move Aegis from a flat vector+metadata model toward a research-grade multi-layered memory architecture.

Motivation

Human cognition uses different memory systems for different purposes. Similarly, AI agents benefit from separating:

What happened (episodic) from what is known (semantic)
How to do things (procedural) from what to avoid (control)

This separation enables more targeted retrieval, better session continuity, and richer agent self-improvement.

Memory Types

Episodic

Time-ordered interaction traces tied to a specific session. Episodic memories capture what happened during agent interactions. Default scope: agent-private (personal interaction trace)

POST /memories/typed/episodic
{
    "content": "User asked about pricing for enterprise plan",
    "agent_id": "sales",
    "session_id": "conv-42",
    "sequence_number": 1
}

Key fields:

session_id (required) — Links the memory to a conversation session
sequence_number (optional) — Explicit ordering within the session

Use when: Recording interaction history, building conversation timelines, debugging agent behavior.

Semantic

Facts, preferences, and knowledge about entities. Semantic memories capture what is known and persist across sessions. Default scope: global (shared knowledge)

POST /memories/typed/semantic
{
    "content": "User is a Python developer based in Manchester",
    "entity_id": "user_123"
}

Key fields:

entity_id (optional) — Links the memory to a specific entity (user, project, concept)

Use when: Storing user profiles, extracted facts, domain knowledge, entity attributes.

Procedural

Workflows, strategies, and reusable patterns. Procedural memories capture how to do things. Default scope: global (reusable strategies)

POST /memories/typed/procedural
{
    "content": "For API pagination, use cursor-based approach",
    "agent_id": "executor",
    "steps": ["Initialize cursor", "Fetch page", "Check has_more flag"],
    "trigger_conditions": ["API returns paginated results"]
}

Key fields:

steps (optional) — Ordered list of steps, stored in metadata
trigger_conditions (optional) — When to apply this procedure, stored in metadata

Use when: Recording successful strategies, building agent playbooks, codifying best practices.

Control

Meta-rules, error patterns, and constraints. Control memories capture what to avoid and behavioral boundaries. Default scope: global (system-wide rules)

POST /memories/typed/control
{
    "content": "Never use range() for unknown-length pagination",
    "agent_id": "reflector",
    "error_pattern": "pagination_incomplete",
    "severity": "high"
}

Key fields:

error_pattern (optional) — Categorizes the error type
severity (optional) — Priority level, stored in metadata
source_trajectory_id (optional) — Links to the trajectory that triggered this rule

Use when: Recording failure modes, enforcing constraints, building guardrails.

Querying Typed Memories

Type-Filtered Search

Search across specific memory types:

POST /memories/typed/query
{
    "query": "pagination strategies",
    "memory_types": ["procedural", "control"],
    "top_k": 10
}

Session Timeline

Get all episodic memories for a session, ordered by sequence:

GET /memories/typed/episodic/session/conv-42

Entity Facts

Get all semantic memories for an entity:

GET /memories/typed/semantic/entity/user_123

Existing Query Enhancement

The existing /memories/query endpoint now accepts a memory_types filter:

POST /memories/query
{
    "query": "user preferences",
    "memory_types": ["semantic", "standard"]
}

Relationship to Existing ACE Types

The 4 new types complement the existing 5 types — they don’t replace them:

Existing Type	New Equivalent	Difference
`standard`	`semantic`	`semantic` adds `entity_id` linking
`strategy`	`procedural`	`procedural` adds structured `steps` and `trigger_conditions`
`reflection`	`control`	`control` adds `severity` and broader meta-rule scope
`progress`	`episodic`	`episodic` adds `session_id` and `sequence_number` for finer granularity
`feature`	—	No equivalent; feature tracking remains unique

Use whichever type best fits your use case. Both sets work side-by-side.

When to Use Which Type

Scenario	Recommended Type
Recording what happened in a conversation	`episodic`
Storing a fact about a user or entity	`semantic`
Saving a reusable workflow or pattern	`procedural`
Recording a rule or constraint	`control`
General-purpose memory storage	`standard`
Storing a lesson from a failure	`reflection` or `control`
Tracking session progress	`progress`
Tracking feature completion	`feature`

Database Schema

Three new columns on the memories table:

Column	Type	Purpose
`session_id`	`String(64)`	Links episodic memories to sessions
`entity_id`	`String(128)`	Links semantic memories to entities
`sequence_number`	`Integer`	Ordering within a session

Two partial indexes for efficient lookups:

ix_memories_session — (project_id, session_id) WHERE session_id IS NOT NULL
ix_memories_entity — (project_id, entity_id) WHERE entity_id IS NOT NULL

Guides

Typed Memory

Typed Memory

Motivation

Memory Types

Episodic

Semantic

Procedural

Control

Querying Typed Memories

Type-Filtered Search

Session Timeline

Entity Facts

Existing Query Enhancement

Relationship to Existing ACE Types

When to Use Which Type

Database Schema

Next Steps

ACE Patterns

API Reference

Guides

​Typed Memory

​Motivation

​Memory Types

​Episodic

​Semantic

​Procedural

​Control

​Querying Typed Memories

​Type-Filtered Search

​Session Timeline

​Entity Facts

​Existing Query Enhancement

​Relationship to Existing ACE Types

​When to Use Which Type

​Database Schema

​Next Steps

ACE Patterns

API Reference

Typed Memory

Motivation

Memory Types

Episodic

Semantic

Procedural

Control

Querying Typed Memories

Type-Filtered Search

Session Timeline

Entity Facts

Existing Query Enhancement

Relationship to Existing ACE Types

When to Use Which Type

Database Schema

Next Steps