Description

APEX-MEM (arXiv:2604.14362, Amazon AGI) is a conversational memory system that combines three innovations relevant to Zeph's MAGMA graph memory:

1. Property graph with domain-agnostic ontology — structures conversations as temporally grounded events in an entity-centric framework
2. Append-only storage — preserves the full temporal evolution of information (no destructive updates)
3. Multi-tool retrieval agent — understands and resolves conflicting or evolving information at query time, producing a compact and contextually relevant memory summary

Results: 88.88% accuracy on LOCOMO QA and 86.2% on LongMemEval — outperforming session-aware state-of-the-art.

Relevance to Zeph

Zeph's MAGMA graph uses typed edges with temporal versioning, but the current belief revision model (tracked in #2441) struggles when varying relation strings defeat conflict detection. APEX-MEM's approach:

• Append-only edges would eliminate the conflict-detection problem: instead of detecting conflicts, each new assertion becomes a new versioned edge, and the retrieval agent resolves the most current truth at query time
• Domain-agnostic ontology aligns with MAGMA's flexible typed-edge model but adds normalization so "works_at" and "employed_by" map to the same relation type
• Multi-tool retrieval for conflict resolution at query time is a direct fit for the SYNAPSE spreading-activation recall — add a conflict resolution pass after recall

User Stories

• As a user, when I update a fact (e.g., job change), I want the agent to surface the most current information, not the earliest version
• As a developer, I want memory graph integrity to not depend on fuzzy string matching for conflict detection

Implementation Sketch

1. Extend MAGMA edge schema with append-only insert semantics (no edge updates — only new versioned edges with supersedes pointer)
2. Add an ontology normalization layer for common relation predicates
3. Add conflict-resolution pass in SYNAPSE recall: when multiple edges assert conflicting values for the same (subject, predicate) pair, use LLM or recency heuristic to select the authoritative value
4. Benchmark on internal recall quality metrics

Source

• Paper: https://arxiv.org/abs/2604.14362
• HTML: https://arxiv.org/html/2604.14362
• Related: research(memory): Kumiho formal belief revision for graph memory — AGM-compliant versioned edge updates (arXiv:2603.17244) #2441 (AGM-compliant belief revision)

Estimated Complexity

Medium. Append-only edge schema is a schema migration. Conflict resolution pass in SYNAPSE is a new module. Ontology normalization is a lookup table initially.

Spec

Pending /sdd specify session.