Landscape Research — 2026

What the ecosystem of knowledge/memory systems looks like, what they teach, and what makes this project's approach unique. Research conducted April 2026.

The Systems

Zep / Graphiti — Temporal Knowledge Graph

Bi-temporal model with four timestamps per edge: creation, expiration, validity start, invalidation. Old facts invalidated rather than deleted.

Why it exists: RAG treats facts as isolated embeddings. "I used to like Adidas, now I prefer Nike" — similarity search returns the Adidas fact because it doesn't know it's superseded. Facts need temporal validity, not just existence.

Key insight: System time (when the system learned) and event time (when it was true) are different questions. Both are needed for audit/compliance.

MAGMA — Multi-Graph Agentic Memory (January 2026)

Four orthogonal graphs over the same event nodes: temporal, causal, semantic, entity. Query intent determines which graph to traverse.

Why it exists: "Why did X happen?" and "When did X happen?" need fundamentally different retrieval. Semantic similarity entangles these.

Key insight: Different query intents need different structural views of the same data.

Mem0 — Production Agent Memory

Four composable scopes (user, agent, run, org). Three memory types (episodic, semantic, procedural). Optimized for production latency.

Why it exists: Memory bleed between agents in production. Without scope isolation, a billing agent retrieves tech support memories.

Key insight: Scopes must compose, not just isolate. Rigid scoping is too inflexible for real queries.

Cognee — Knowledge Engine

Universal "DataPoint" primitive. Triple-store hybrid. "Memify" step prunes stale nodes and strengthens frequent connections.

Why it exists: Knowledge graphs degrade without maintenance. Stale nodes accumulate. Re-indexing takes days and breaks.

Key insight: Memory is not static storage — it's an evolving structure that must adapt.

LightRAG — Fast RAG Pipeline

Entities + relations extracted by LLM, each with textual profiles + embeddings. Dual-level retrieval: entity-specific vs thematic. ~100x cheaper than Microsoft GraphRAG.

Why it exists: GraphRAG proved graph structure improves retrieval but is catastrophically expensive (610K tokens for community detection).

Key insight: Specific-entity queries and thematic queries need different retrieval strategies.

Letta / MemGPT — OS-Like Memory Hierarchy

Memory as tiers: core (in-context), recall (history), archival (indexed). The agent manages its own memory promotion/demotion.

Why it exists: Fixed context windows. Old conversations lost, documents that exceed the window can't be analyzed.

Key insight: Context pollution is worse than missing context. Too much irrelevant information actively degrades performance.

Cross-Cutting Findings

What's Unique About This Project

No system in the landscape combines version control with knowledge structure. Atomic commits, branching, and DAG history come from git's lineage, not from AI/ML research. This is a unique design space.

No system uses set-intersection scoping. Every system uses entity-relation triplets with vector embeddings as primary retrieval. The dimensional/scoping model — where navigation is set intersection over placements — is not found elsewhere.

Structure as retrieval. All other systems use embeddings as primary or co-primary retrieval. This project's scope operation is purely structural — no cosine similarity, no vector search.

Reader-determined meaning. Other systems bake meaning at write time (LLM-generated entity profiles, relation summaries). This project leaves meaning in the content. The structure tells the reader where to look; the reader discovers what it means.

The Triplets vs Scoping Question

The exploration tested whether entity-relation triplets (the dominant paradigm) or dimensional scoping (this project's model) is the right primitive. 18 agents across three rounds of exploration.

Findings:

• Three agents designing from scratch all converged on directed relationships (graph primitives)
• Three scenario tests all found the same gaps in the pure dimensional model (no direction, no entity identity, no traversal)
• One critical reviewer challenged the test scenarios as biased toward graph-native domains

Resolution: The false binary was rejected. The converged design uses one primitive (chunk) where connections are chunks placed on multiple scopes — preserving set-intersection scoping while enabling traversal through the placement graph. Direction was identified as needed but may emerge from scope context rather than requiring explicit structural direction.

What Others Have That This Project Addresses Differently