Zep: A Temporal Knowledge Graph Architecture for Agent Memory

TL;DR

Introduces Zep, a production “memory layer service” for agents built on Graphiti, a temporally-aware knowledge graph.
Uses a three-tier graph: episodes (raw, non-lossy inputs), semantic entities/facts (derived relations with validity), and communities (clusters with high-level summaries).
Key mechanism: a bi-temporal model that distinguishes event time (when a fact is true) from transaction time (when Zep learned/invalidated it), enabling historical queries and update semantics.
Retrieval is an explicit 3-stage pipeline: search → rerank → construct context, mixing cosine similarity, BM25, and graph BFS.
Reports strong results on DMR and LongMemEval, with large reductions in context tokens and latency vs full-context baselines (as reported).

A concrete operationalization of temporal validity intervals (tvalid, tinvalid) for facts, plus separate audit timestamps (t'created, t'expired).
A non-lossy episodic store linked bidirectionally to derived semantic edges/entities for provenance (“trace a fact back to its source episode”).
A hierarchical “episodes → entities/facts → communities” graph organization that mirrors common agent memory desiderata (episodic + semantic + higher-level consolidation).
Treats retrieval as a configurable IR stack (hybrid search + reranking), not “just query a graph”.

Episodes: raw messages/text/JSON with a reference timestamp tref.
Entities: extracted/resolved nodes with summaries and embeddings.
Facts / semantic edges: relations between entities with temporal metadata; can be invalidated by newer contradictory edges.
Communities: clusters of entities with summaries for higher-level retrieval/global context.

Write path:
- Ingest episode → extract entities (with short context window) → entity resolution/dedup → extract facts/edges → temporal extraction → invalidate overlapping contradictions.
- Community detection + summary updates (dynamic extension + periodic refresh).
Read path:
- Query → search candidates (cosine/BM25/BFS) → rerank (RRF/MMR/mention- and distance-based; optional cross-encoder) → construct a context string with facts + validity ranges and entity summaries.
Maintenance:
- Continuous dedup/resolution; temporal invalidation rather than deletion to preserve history.
- Community refreshes are an explicit long-running maintenance job.

DMR (MemGPT’s Deep Memory Retrieval task):
- Reports Zep outperforming MemGPT slightly (e.g., 94.8% vs 93.4% with gpt-4-turbo; full-context is also high).
- Also reports gpt-4o-mini runs (Zep 98.2% vs full-context 98.0%).
LongMemEval (LongMemEvals subset):
- Conversations average ~115k tokens.
- Reports accuracy gains vs full-context baselines while reducing average context tokens to ~1.6k and reducing latency by ~90%.
- Breakdown shows improvements are strongest for preference/multi-session/temporal questions, with a notable drop on “single-session-assistant”.

Validity intervals + invalidation semantics as the default for “corrections” (instead of overwriting/deleting).
Hybrid retrieval that includes graph BFS expansion (seeded by initial search) to surface contextually related facts.
Explicit constructor step that formats memory into a stable prompt-ready layout (facts with date ranges, entity summaries).
Storing provenance links from derived facts → source episodes to support citation/audit.

Cost profile of graph construction + community summarization at scale (background jobs, LLM calls) is only partially characterized.
Security/poisoning is not the focus; a memory layer service needs explicit tenant isolation + write-time policy gates.
DMR is acknowledged as a weak benchmark; LongMemEval baselines are stronger but still rely on LLM-judge evaluation.