Product Deep-Dive

MNEMON

MNEMON (Managed Neural Engine for Memory Organization & Navigation)

The contextual memory engine powering both OAR and AIEGES. Three memory tiers, a dual-graph architecture, and a ritual-based review process that ensures AI agents never lose context.

Three-Tier Memory Architecture

Hot Memory

Active working context. Redis-backed with sub-millisecond access. Holds the current conversation, recent tool outputs, and in-flight planning state.

Capacity

~50 MB

Latency

<1ms

Persistence

Session-scoped

Cold Memory

Persistent knowledge graph. Stored in Neo4j via Graphiti, enabling temporal queries and semantic search across past sessions.

Capacity

Unlimited

Latency

~5ms

Persistence

Permanent

Federated Memory

Shared knowledge across team members via the Airlock Protocol. Cryptographically verified, selectively replicated, and access-controlled.

Capacity

Policy-bound

Latency

~20ms

Persistence

Replicated

Context-Slicing Protocol

Context windows are finite. MNEMON solves this by slicing knowledge into ~2000-token segments, each tagged with semantic metadata and group IDs for efficient retrieval.

┌──────────────────────────────────────┐
│         Full Context (128K tokens)   │
├──────────────────────────────────────┤
│  Slice 1 (~2000 tokens)             │
│  ├─ group: "deployment-pipeline"    │
│  ├─ archetype: procedural           │
│  └─ relevance: 0.94                 │
├──────────────────────────────────────┤
│  Slice 2 (~2000 tokens)             │
│  ├─ group: "auth-service"           │
│  ├─ archetype: semantic             │
│  └─ relevance: 0.87                 │
├──────────────────────────────────────┤
│  Slice N ...                         │
│  Only slices above threshold are     │
│  loaded into the active context.     │
└──────────────────────────────────────┘

Dual-Graph Architecture

MNEMON combines two graph systems for comprehensive memory management:

Graphiti (Knowledge Graph)

Temporal knowledge graph backed by Neo4j. Stores entities, relationships, and their evolution over time. Enables queries like “What changed since last Tuesday?”

  • • Temporal edges with valid-time semantics
  • • Semantic similarity search via embeddings
  • • Entity resolution and deduplication
  • • Bi-directional relationship traversal

LangGraph (Orchestration Graph)

Stateful workflow graph that orchestrates memory operations. Manages the retrieval, slicing, scoring, and injection pipeline.

  • • Stateful nodes with checkpointing
  • • Conditional branching based on relevance scores
  • • Parallel retrieval across memory tiers
  • • Built-in retry and fallback paths

Menoms Review Ritual

Every memory write goes through a four-step synthesis process called the Menoms Review:

1

Extract

Identify key entities, relationships, and facts from the raw input.

2

Classify

Assign a memory archetype (Episodic, Semantic, Procedural, Checkpoint, Hybrid) and group ID.

3

Reconcile

Compare against existing knowledge. Resolve conflicts, merge duplicates, update temporal edges.

4

Commit

Write the reconciled memory to the appropriate tier with a full audit trail.

5 Memory Archetypes

E

Episodic

Timestamped event sequences. 'What happened during the last deployment?'

S

Semantic

Conceptual relationships. 'How does component X relate to service Y?'

P

Procedural

Step-by-step workflows. 'How do we run the release pipeline?'

C

Checkpoint

Snapshot-in-time state. 'Restore the project state from Tuesday.'

H

Hybrid

Cross-archetype queries. 'What happened (episodic) and why does it matter (semantic)?'

Explore the full AIEGES stack

MNEMON powers the memory layer for both OAR and AIEGES.

Smart Gateway →AIEGES Quickstart →