● NODE-01 ONLINE AIR_GAP ENFORCED EGRESS 0 B PIPELINE V5.3 27B OPEN_WEIGHT // LOCAL
Local_SOTA Pipeline_Online

AI_Memory_Isn't_A_Model_Problem.
It's_An_Engineering_Problem.
We_Have_The_Numbers.

AIR-GAPPED Zero_Cloud_Dependency · Sub-$2K_Hardware

Air-gapped memory retrieval. Thousands of sessions. Zero cloud dependency. Echo Locate runs a full index → filter → rank → synthesize pipeline on consumer hardware, matching cloud-scale systems on standardized benchmarks while keeping every byte local — on under $2,000 of consumer hardware.

Request_Pilot ›› Tech_White_Paper
#1
LongMemEval_Score
Locally_Hosted_Leaderboard
#6
LongMemEval_Score
Overall_Leaderboard
LongMemEval-S · GPT-4o_JudgedVerified
85.40%
QA_Score · Single_Pass
#1
Overall — Locally_Hosted_Board
#6
Overall — Full_Leaderboard
Reader: QWEN3.6_27B Compute: Local Cloud_Dep: 0.0%
Secure_Domains

Secure Domains

Legal · Healthcare · Defense · Quant_Finance
Legal_Discovery
Privileged / E-Discovery

Summarize, redact, and cite millions of privileged documents inside the firm's enclave.

Chain_of_Custody
Healthcare_PHI
HIPAA / HITECH / EMA

PHI-safe diagnostics and molecular search running beside the MRI, never to a tenant.

PHI_Contained
Defense_Intel
DOD / ITAR / IL6

Classified inference for tactical edge, autonomy loops, and sensor fusion under air-gap.

Cleared_Deploy
Quant_Finance
SEC / FINRA / SOC2-II

Low-latency signal engines and risk reasoners that never leave the trading enclave.

Audit_Ready
LongMemEval_Benchmark_Analysis

LongMemEval Benchmark_Analysis

Quantized_27B_models_on_decade-old_equipment_breaking_into_the_leaderboard
Officially_GPT-4o_judged · Single_pass_QA · Verified_leaderboard · 16_entries
Leaderboard — LongMemEval-S GPT-4o · T=0
#System / RunReader_ModelCompute QA_Score · 4o_judged
01MASTRA_OMGPT-5-MINICloud
94.87%
02MASTRA_OMGEMINI-3-PRO-PREVIEWCloud
93.27%
03HINDSIGHTGEMINI-3-PRO-PREVIEWCloud
91.40%
04MASTRA_OMGEMINI-3-FLASH-PREVIEWCloud
89.20%
05HINDSIGHTGPT-OSS-120BCloud
89.00%
06ECHO_LOCATEQWEN_3.6_27BLocal
85.40%
07SUPERMEMORYGEMINI-3-PRO-PREVIEWCloud
85.20%
08SUPERMEMORYGPT-5Cloud
84.60%
09MASTRA_OMGPT-4oCloud
84.23%
10HINDSIGHTGPT-OSS-20BCloud
83.60%
T-11EMERGENCEMEM_SIMPLEGPT-4oCloud
82.40%
T-11ECHO_LOCATEQWEN_3.5_27BLocal
82.40%
13SUPERMEMORYGPT-4oCloud
81.60%
14MASTRA_RAGGPT-4oCloud
80.05%
15ZEP_GRAPHITIGPT-4oCloud
71.20%
16FULL_CONTEXTGPT-4oCloud
60.20%
Local / Air_Gapped Cloud / 3P_Hosted
16_Verified_Entries · GPT-4o_Judged · Single_Pass_QA
Footnote — EmergenceMem Internal (86.00%) is intentionally excluded from this leaderboard. Mastra's research page documents that this configuration is "not publicly reproducible." We exclude unreproducible results so every entry on this board can be independently verified. EmergenceMem's reproducible Simple configuration is included at 82.40%. SRC: mastra.ai/research/observational-memory · supermemory.ai/research
Per_Category_Head_to_Head

Per_Category_ Head_To_Head

Breakdown_against_the_top_cloud_hosted_systems
8_Systems · 6_Categories · Sorted_high_to_low_within_category · GPT-4o_judged
Echo_Locate Cloud_Competitors
CAT_01

Temporal_Reasoning

01
MASTRA_OM · GPT-5-MINI
95.50%
02
ECHO_LOCATE · QWEN_3.6_27B
92.48%
03
MASTRA_OM · GPT-4o
85.70%
04
SUPERMEMORY · GEMINI-3-PRO
81.95%
05
SUPERMEMORY · GPT-5
81.20%
06
SUPERMEMORY · GPT-4o
76.69%
07
ZEP
62.40%
08
FULL_CONTEXT
45.10%
CAT_02

Knowledge_Update

01
MASTRA_OM · GPT-5-MINI
96.20%
T-02
ECHO_LOCATE · QWEN_3.6_27B
89.74%
T-02
SUPERMEMORY · GEMINI-3-PRO
89.74%
04
SUPERMEMORY · GPT-4o
88.46%
05
SUPERMEMORY · GPT-5
87.18%
06
MASTRA_OM · GPT-4o
85.90%
07
ZEP
83.30%
08
FULL_CONTEXT
78.20%
CAT_03

Multi_Session

01
MASTRA_OM · GPT-5-MINI
87.20%
02
MASTRA_OM · GPT-4o
79.70%
03
SUPERMEMORY · GEMINI-3-PRO
76.69%
04
ECHO_LOCATE · QWEN_3.6_27B
75.94%
05
SUPERMEMORY · GPT-5
75.19%
06
SUPERMEMORY · GPT-4o
71.43%
07
ZEP
57.90%
08
FULL_CONTEXT
44.30%
CAT_04

Single_Session_User

01
MASTRA_OM · GPT-4o
98.60%
02
SUPERMEMORY · GEMINI-3-PRO
98.57%
T-03
SUPERMEMORY · GPT-5
97.14%
T-03
SUPERMEMORY · GPT-4o
97.14%
05
ECHO_LOCATE · QWEN_3.6_27B
95.71%
06
MASTRA_OM · GPT-5-MINI
95.70%
07
ZEP
92.90%
08
FULL_CONTEXT
81.40%
CAT_05

Single_Session_Assistant

01
SUPERMEMORY · GPT-5
100.00%
02
SUPERMEMORY · GEMINI-3-PRO
98.21%
03
SUPERMEMORY · GPT-4o
96.43%
T-04
MASTRA_OM · GPT-5-MINI
94.60%
T-04
FULL_CONTEXT
94.60%
06
ECHO_LOCATE · QWEN_3.6_27B
85.71%
07
MASTRA_OM · GPT-4o
82.10%
08
ZEP
80.40%
CAT_06

Single_Session_Preference

01
MASTRA_OM · GPT-5-MINI
100.00%
02
SUPERMEMORY · GPT-5
76.67%
03
MASTRA_OM · GPT-4o
73.30%
T-04
SUPERMEMORY · GEMINI-3-PRO
70.00%
T-04
SUPERMEMORY · GPT-4o
70.00%
06
ECHO_LOCATE · QWEN_3.6_27B
60.00%
07
ZEP
56.70%
08
FULL_CONTEXT
20.00%
Echo's_Neural_Pipeline

Index → Filter → Rank → Synthesize

Topographic_map · Model_size_27B · Open_weight · Local_hosted
Path: /Core/Cortex/Schema · Sys: Neural_Network_V5.3
// Activity_Feed [Live]
03.05.2026ECHO_LOCATE_ALPHA — 71% — FIRST_COMPLETE_RUN
04.03.2026ECHO_LOCATE — 81.0% — SECOND_COMPLETE_RUN
04.19.2026ECHO_LOCATE_V8.x.1 — 82.4% — QWEN3.5_GENERATOR
04.25.2026ECHO_LOCATE_V8.x.1 — 85.40% — QWEN3.6_Q4_K_XL_SOTA
// Generation DUAL_INSTANCES
ModelQWEN3.6-27B
QuantQ4_K_XL
Size_on_Disk~17 GB
VRAM_Loaded~21.5 GB
Context_KV262K
RTX_3090 ×2 HP_Z840
// Embedding SHARED
ModelQWEN3-EMBEDDING-4B
QuantQ8_8BIT
Size_on_Disk~4.1 GB
VRAM_Loaded~4.8 GB
RuntimeLocal
RTX_5060_Ti ThinkStation_S20
// Reranker SHARED
ModelQWEN3-RERANKER-4B
QuantFP16_FULL
Loaded_ViaAutoModelForCausalLM
Size_on_Disk~8 GB
VRAM_Loaded~8.1 GB
RTX_5060_Ti ThinkStation_S20
$2K_Node_Spec

$2K Node_Spec

Deployed_chassis // field_units // scrap_workstations_vs_frontier_cloud
Chassis_Info 2_Field_Units
HP Z840 workstation tower
Gen_Node // Built_2014
HP_Z840
2× RTX_3090 · QWEN3.6_27B_Q4_K_XL_DUAL_INSTANCES 2× XEON_E5-2650_V4 64GB_DDR4_RAM 512GB_NVME Ubuntu
Lenovo ThinkStation S20 workstation tower
Embed+Rerank // Built_2009
ThinkStation_S20
RTX_5060_Ti · EMBED_4B + RERANK_4B 1× INTEL_XEON_W3680 16GB_DDR3_RAM 128GB_SATA_SSD Ubuntu
System_Log // node-01 tail -f
08:39:02.114 kernel: [memory.subsystem] index store mounted at /core/cortex/idx — 2,411,872 vectors
08:39:02.119 kernel: [embed.qwen3-4b] loaded Q8_8BIT — 4.8 GB VRAM — OK
08:39:02.668 kernel: [rerank.qwen3-4b] loaded FP16 — 8.1 GB VRAM — OK
08:39:03.041 kernel: [gen.qwen3.6-27b] dual instance warmup — 38.2 tok/s — OK
08:39:03.412 kernel: [net.airgap] egress interface disabled — ENFORCED
08:39:04.002 kernel: [pipeline] session #8841 — index → filter (n=1,204) → rank (k=32)
08:39:04.187 kernel: [pipeline] session #8841 — synthesize — 212 tokens — 5.6 s
08:39:05.224 kernel: [pipeline] session #8842 — index → filter (n=987) → rank (k=28)
08:39:05.501 kernel: [audit] retrieval trace sealed — sha256:9f2c…e41a
08:39:06.033 kernel: [thermal] core 78 °C — fan curve nominal
08:39:06.612 kernel: [pipeline] session #8842 — synthesize — 184 tokens — 4.8 s
08:39:07.140 kernel: [kv] context window 262K — utilization 61%
08:39:07.955 kernel: [net.airgap] 0 bytes egress — CLEAN
08:39:08.221 kernel: [pipeline] session #8843 — index → filter (n=1,311) → rank (k=32)
08:39:08.870 kernel: [audit] retrieval trace sealed — sha256:41bd…0c77
08:39:09.412 kernel: [embed.qwen3-4b] batch 128 → 0.9 s — OK
08:39:10.001 kernel: [gen.qwen3.6-27b] instance B active — load balanced
08:39:10.518 kernel: [pipeline] session #8843 — synthesize — 246 tokens — 6.4 s
08:39:11.209 kernel: [thermal] core 78 °C — stable
08:39:12.077 kernel: [net.airgap] 0 bytes egress — CLEAN
Pilot_Enrollment

Pilot Enrollment

Interested in learning more? Please share your use case and contact info below.

Pilot_Request // 24H_ETA
› Pilot_Request_Received // ETA 24H