/ ml vs Hindsight

A frontier AI memory system vs. multi-pass agentic reflect loops across four memory tables.

/ ml Frontier AI Memory System
  • ✓94.5% LoCoMo (+8.9 pts vs OSS / +4.9 pts vs Gemini-3) & 80.8% DolphinBench (+11.3 pts)
  • ✓0.86 ms P50 recall & 33.9s median task latency (zero multi-minute reflection stalls)
  • ✓$45.90 DolphinBench run cost (46% cheaper) & scales to 88.9% (100K) / 77.2% (10M) on BEAM
Hindsight Hindsight (Vectorize) 4-Network LLM Retain + Agentic Reflect
  • ✕85.6% LoCoMo on OSS-120B (reaches 89.6% only by swapping judge to Gemini-3 Pro) & 69.5% DolphinBench
  • ✕1,500+ ms recall & 158.9s–602.9s P95 tail latency from query-time agentic reflect loops
  • ✕$84.65 DolphinBench run cost (nearly 2× higher) & scores only 66.2% at 100K on BEAM
LoCoMo Accuracy +8.9 pts
/ ml 94.5%
Hindsight Hindsight 85.6%
89.6% when Hindsight swaps to Gemini-3 Pro
DolphinBench (600) +11.3 pts
/ ml 80.8%
Hindsight Hindsight 69.5%
600-task multi-session agent leaderboard
Recall & Tail Latency 1,744× Faster
/ ml 0.86 ms
Hindsight Hindsight 1,500 ms
33.9s vs 158.9s P95 task tail latency
Dolphin 600 Cost -46% Cost
/ ml $45.90
Hindsight Hindsight $84.65
Zero query-time reflect loop token burn

How / ml outperforms Hindsight across accuracy, tail latency, and cost

Evaluated across LoCoMo (ACL 2024; Hindsight arXiv:2512.12818), the independent 600-task DolphinBench Official Leaderboard, and BEAM (ICLR 2026).

LoCoMo · DolphinBench · BEAM · Higher is Better

Memory & Agent Task Accuracy

+8.9 pts LoCoMo · +11.3 pts Dolphin
/ ml LoCoMo Top-20
94.5%
Hindsight Gemini-3 Pro Swap
89.6%
Hindsight Paper OSS 120B
85.6%
Hindsight Paper OSS 20B
83.6%
/ ml DolphinBench (600)
80.8%
Hindsight DolphinBench (600)
69.5%
/ ml BEAM 100K → 10M
88.9%
Hindsight BEAM 100K
66.2%
Recall P50 · Task P50/P95 · Run Cost · Lower is Better

Latency, Tail Spikes & Run Cost

0.86 ms vs 1,500+ ms · No 159s Spikes
/ ml Recall P50
0.86 ms
Hindsight Recall P50
1,500 ms
/ ml Dolphin Task P50
33.9 s
Hindsight Dolphin Task P50
55.3 s
Hindsight Dolphin Tail P95
158.9 s
/ ml Dolphin 600 Cost
$45.90
Hindsight Dolphin 600 Cost
$84.65

Three structural bottlenecks behind multi-pass agentic reflection loops

Hindsight (Vectorize, arXiv:2512.12818) splits memory into four logical networks and runs iterative LLM reflection loops (TEMPR + CARA). Here is why that hurts production agents—and why / ml wins.

Multi-pass agentic reflect loops stall interactive agents

/ ml 0.86 ms Recall · 33.9s Task P50
Hindsight Hindsight 1,500+ ms Recall · 158.9s Task P95

Hindsight's reflect pipeline executes multi-pass LLM reasoning loops over retrieved memories at query time. On the 600-task DolphinBench leaderboard, this pushed Hindsight's median task time to 55.3 seconds and P95 tail latency to 158.9–602.9 seconds (up to 10 minutes).

/ ml Advantage: Resolves multi-hop state in 0.86 ms (33.9s task P50) with zero query-time reflection loops.

85.6% on standard LoCoMo—reaching 89.6% only by swapping the judge LLM

/ ml 94.5% SOTA (Top-20 Retrieval)
Hindsight Hindsight 85.6% OSS-120B (89.6% Gemini-3 Swap)

In Hindsight's paper (arXiv:2512.12818), their architecture scores 83.6% (20B) and 85.6% (120B) on LoCoMo under standard evaluation, reaching their advertised 89.6% figure only by swapping the answering backbone model to Gemini-3 Pro alongside multi-pass reflection.

/ ml Advantage: Achieves 94.5% SOTA at strict Top-20 Retrieval—beating Hindsight's standard score by +8.9 pts and even beating their Gemini-3 Pro run by +4.9 pts.

Four-network LLM extraction burns nearly 2× the budget for lower accuracy

/ ml 80.8% Dolphin ($45.90) · 88.9% BEAM
Hindsight Hindsight 69.5% Dolphin ($84.65) · 66.2% BEAM

Maintaining four separate memory stores (world facts, agent experiences, entity summaries, and evolving beliefs) requires heavy LLM extraction on every retain call. Yet on DolphinBench (600 tasks), Hindsight scores just 69.5% at $84.65, and scores 66.2% at 100K on BEAM.

/ ml Advantage: Delivers 80.8% on DolphinBench (+11.3 pts) at $45.90 (-46% cost), and 88.9% at 100K / 77.2% at 10M on BEAM (+22.7 pts).

Complete Head-to-Head Breakdown: / ml vs. Hindsight

Every benchmark, latency percentile, and billing line-item side by side.

Dimension Hindsight (Vectorize) / ml Your Advantage with / ml
Core Architecture 4-Network LLM Retain + Multi-Pass Agentic Reflect Frontier AI Memory System Single synchronous state machine; zero LLM reflection loops
LoCoMo Accuracy 83.6% (20B) · 85.6% (120B) · 89.6% (Gemini-3 Pro) 94.5% SOTA (Top-20 Retrieval) +8.9 pts over Hindsight OSS (+4.9 pts over Gemini-3 Pro)
DolphinBench (600 Tasks) 69.5% accuracy · $84.65 run cost 80.8% accuracy · $45.90 run cost +11.3 pts higher accuracy at 46% lower total cost
BEAM Long-Horizon Benchmark 66.2% (at 100K tokens) 88.9% (100K) · 83.5% (1M) · 77.2% (10M) +22.7 pts at 100K and scales cleanly to 10M tokens
Recall Latency (P50) 1,500+ ms (TEMPR + CARA reflection) 0.86 ms 1,744× faster recall (sub-millisecond)
DolphinBench Task Latency (P50 / P95) 55.3 s P50 · 158.9 s P95 (602.9 s on M3) 33.9 s P50 39% faster median task; eliminates 2.6–10 min tail spikes
Write Speed & Compute LLM extraction across 4 networks on every retain() Synchronous < 10 ms · $0 external LLM API fees Instant writes with zero internal LLM extraction calls
Codebase & Symbol Graphs — Textual world/belief networks only Full AST Symbol, Call-Graph & Bug Recall Built for both conversational agents and coding harnesses

Replace Hindsight retain & reflect with / ml

Replace slow multi-pass reflect() loops with synchronous .add() and 0.86 ms .search().

1

Swap Hindsight retain/reflect for 0.86 ms Stateful Recall

Replace Hindsight's retain() and multi-second reflect() calls with / ml's synchronous .add() and .search().

agent.py — Hindsight to / ml Migration Diff Python / TypeScript 
-from hindsight_client import Hindsight-client = Hindsight(base_url="http://localhost:8888")+from slashml import Memory+mem = Memory(api_key="<SLASHML_API_KEY>") -# Hindsight runs 4-network LLM extraction & slow multi-pass reflect() loops-client.retain(bank_id="agent_01", content=fact)-answer = client.reflect(bank_id="agent_01", query=question)+# / ml updates state in milliseconds and recalls multi-hop state in 0.86 ms:+mem.add(fact, user_id="agent_01")+memories = mem.search(question, user_id="agent_01")
2

Import Existing Hindsight Memory Banks

Import your existing Hindsight memory banks into / ml with instant synchronous indexing and zero external LLM fees.

Import Hindsight Banks into / ml CLI Migration 
$# Export & ingest existing Hindsight memory banks into / ml:+npx -y @slashml/memory migrate --from hindsight --bank-url http://localhost:8888 --api-key <SLASHML_API_KEY>

Ready to upgrade from Hindsight in 60 seconds?

Start free with 10,000 ingests + 1,000 retrievals/month and zero feature locks, or contact our team for custom VPC migration support.

Start Free & Get API Key Contact Engineering Support