Comparison

Tenure vs. Zep

Zep is a temporal knowledge graph for enterprise agent memory. Multi-container deployment. 897 seconds to ingest 35 beliefs. Mean precision: 0.09.

TL;DR

  • Zep's graph-based retrieval achieves 0.09 mean precision. Zero active retrieval passes.
  • Ingestion cost: 897 seconds for 35 beliefs (mean 25,630ms per belief).
  • A belief introduced at turn 1 may not be queryable until the session has concluded.
  • Drift score 0.89 on re-entry: noise beliefs from unrelated turns contaminate retrieval.
  • Requires multi-container deployment with graph store and embedding pipeline.
Tenure precision
1.00
43/43 active passes
Zep precision
0.09
0/43 active passes
Tenure latency
9.77ms
p50 retrieval
Zep latency
124.36ms
p50 retrieval

The availability gap

Zep's async ingestion architecture creates a structural availability problem in conversational use. Individual beliefs reach 125,148ms during ingestion. At a mean of 25,630ms per belief against a conversational turn cadence of 10 to 30 seconds, a belief introduced at turn 1 may not be queryable until the session has largely concluded.

Tenure extracts beliefs during the stream and makes them available on the very next turn. The session-turn evaluation confirms this: all 12 session cases pass, including turn N+1 availability.

Memory that takes 15 minutes to become queryable is not memory. It is a batch pipeline. Conversational memory requires sub-second write-to-read latency. Zep's graph construction overhead makes this architecturally impossible at conversational cadence.

Session-level noise isolation

After topic drift, Zep produces drift score 0.89 on the re-entry turn. It surfaces noise beliefs introduced during unrelated drift turns while retaining the correct belief in the result set at low rank. The model receives noise alongside signal and must sort it out at inference time.

Session-turn latency: 418.13ms p50 (3x degradation over single-turn baseline of 139.64ms).

Infrastructure comparison

Zep requires a multi-container deployment including a graph store and embedding pipeline. Tenure runs as a single Docker container on localhost.

Property Tenure Zep
Mean retrieval precision 1.00 0.09
Active retrieval passes 43/43 0/43
Total passes (89 cases) 89/89 9/89
Retrieval latency (p50) 9.77ms 124.36ms
Session latency (p50) 47.79ms 418.13ms
Drift score 0.00 0.89
Ingestion (35 beliefs) 1.0s 897s
Deployment Single container Multi-container (graph + embed)
Runs fully local Always Cloud-first (self-host available)
Account required No Yes
Scope isolation Hard filter None native
Supersession handling Chain with audit Temporal invalidation
Turn N+1 availability Yes No (async ingestion)
Per-turn injection audit Yes No
License MIT Proprietary (cloud) / BSL

Zep evaluated against pinned Docker image digest zep-zep@sha256:f9e5372caf60. Full methodology: arXiv:2605.11325. Dataset: HuggingFace.

One container. Sub-15ms. Precision 1.0.

No graph store. No embedding pipeline. No async ingestion gap. Thirty seconds to install.