Memory Architecture

Squad’s memory architecture is modelled on the structure of human cognition. Rather than treating all knowledge as a single undifferentiated store, Squad maintains orthogonal memory systems: each responsible for a different type of knowledge, with different persistence characteristics and learning rates.

This architecture, called SOMA (Squad Ontology & Memory Architecture), sits between data ingestion and reasoning: providing the persistent knowledge layer that the intelligence engine queries, updates, and learns from.

Orthogonal Memory Systems

Squad maintains four distinct memory types, each inspired by a different brain system:

Memory System	Brain Analogue	What It Stores	Learning Rate
Episodic	Hippocampus	Specific events with full context: document sections, conversation turns, reasoning traces	Per session
Semantic	Temporal/Parietal cortex	General facts and knowledge: entities, relationships, corrected beliefs	Across sessions (requires repetition)
Procedural	Basal ganglia	Skills and action sequences: stored workflows, tool definitions, query templates	Very slow (requires human approval)
Working	Prefrontal cortex	Active task state: current plan, tool calls, search history, conversation context	Instant (session-scoped)

These are not a hierarchy: they are orthogonal dimensions. A single interaction can write to multiple memory systems simultaneously, and the intelligence engine queries across all of them when processing a request.

SOMA

Consolidation flows from fast to slow. The most trusted knowledge takes the longest to form

Instant

Human approval

Prefrontal Cortex

Working

Active task state

Current plan Tool results Search history Conversation

Session-scoped · Ephemeral

Traces
recorded

Hippocampus

Episodic

Events with full context

Document sections Conversations Reasoning traces Decisions

Per session · Persistent

Patterns
crystallise

Temporal / Parietal

Semantic

General facts & knowledge

Entities Relationships Domain concepts Co-occurrence

Across sessions · Repetition

Human
approval

Basal Ganglia

Procedural

Skills & action sequences

Workflows Tool definitions Query templates Compiled procs

Very slow · Human-governed

Semantic Layer

Retrieval Graph

Workflow Graph

Memory Consolidation

Knowledge flows between memory systems through a consolidation pipeline that mirrors how the brain converts short-term experiences into long-term knowledge:

1. Working memory captures the immediate context of a task: the current plan, intermediate results, and conversation state. This is ephemeral and session-scoped.

2. Episodic memory records what happened: full reasoning traces with decision points, tool calls, results, and outcomes. At the end of each session, relevant working memory is consolidated into episodic traces.

3. Semantic memory crystallises general knowledge from repeated patterns across episodes. When the same entities, relationships, or facts appear across multiple interactions, they are promoted into the persistent knowledge graph.

4. Procedural memory captures proven workflows and action sequences. This is the slowest-learning system: a workflow must be explicitly approved through human-in-the-loop review before it becomes a reusable workflow.

Tip

Each memory system has a different learning rate by design. Working memory is instant but ephemeral. Procedural memory takes the longest to form but represents the highest-confidence knowledge: proven patterns that have been validated by a human operator.

Three Graph Domains

Squad’s memory systems are physically organised into three graph domains within a labeled property graph database:

Semantic Layer

Entities, relationships, and ontology types extracted from your data. The structured representation of what Squad knows about your domain.

Explore the Semantic Layer →

Retrieval Graph

Co-occurrence concept networks and community structures that power Squad’s retrieval capabilities: from precise local lookups to broad thematic queries.

Learn about retrieval →

Workflow Graph

The physical store for procedural memory: approved plans, steps, tools, and query templates. The system’s learned repertoire of how to accomplish tasks, built up through human-approved workflows.

See how workflows work →

Fuzzy-Trace Storage

Squad’s storage model is inspired by Fuzzy-Trace Theory from cognitive science: the finding that the brain stores both a verbatim trace (exact details, fragile) and a gist trace (meaning, durable) of every experience.

This maps to Squad’s three storage layers:

Layer	Trace Type	What’s Stored	Deduplication
Layer 0	Verbatim	Raw document segments with full source provenance	None: every instance preserved
Layer 1	Early gist	Extracted mentions, patterns, and relationships anchored in L0	Confidence-scored
Layer 2	Durable gist	Canonical, deduplicated entities resolved across all sources	Full cross-source resolution

This layered approach means Squad can always trace an answer back to its source material (via L0), while still operating efficiently over consolidated, deduplicated knowledge (via L2).

Next Steps

Semantic Layer

How Squad extracts and organises entities, relationships, and domain ontologies from your data.

Read about the Semantic Layer →

Retrieval Graph

How Squad retrieves relevant context using co-occurrence networks and community-aware search.

Read about retrieval →