Multi-Tenant Edge Agent Memory on RK3588 ARM64: O13 Evidence-Boundary Replay in LongMemOS (~3.1 MB)
Short Answer: Running multi-store retail or factory AI agents on a cloud vector database leaks cross-tenant context and drops exact batch numbers when lossy summaries overwrite prior turns. Deploying LongMemOS (~3.1 MB s
Short Answer: Running multi-store retail or factory AI agents on a cloud vector database leaks cross-tenant context and drops exact batch numbers when lossy summaries overwrite prior turns. Deploying LongMemOS (~3.1 MB standalone binary, embedded_models=false) directly on an RK3588 ARM64 edge gateway enforces strict per-store tenant isolation while preserving verbatim source lines via O13 Evidence-Boundary Replay.
Why Do Cloud Vector Databases Fail Multi-Store Edge Agents?
Cloud vector pipelines fail multi-store edge workloads because dense embeddings blur single-digit SKU codes and ignore hard tenant boundaries during approximate nearest-neighbor retrieval. When a regional retail chain operates 30 to 80 franchised locationsβeach with its own shelf-life thresholds, vendor rebates, and local cold-chain logsβshipping raw operational telemetry to a shared hosted embedding index introduces three operational hazards:
- Cross-store metadata bleed: A single missing filter predicate in a shared vector collection lets Store #042 retrieve pricing overrides or supplier defect notes belonging to Store #019.
-
Numeric drift under cosine similarity: Dense embeddings treat batch code
LOT-20261108-AandLOT-20261109-Aas nearly identical vectors, causing expiry-alert agents to cite the wrong warehouse pallet. - Unpredictable WAN latency and token overhead: Edge gateways in basement stockrooms or industrial workshops frequently experience cellular packet drops, stalling remote embedding calls and bloating recurring API bills.
Because LongMemOS runs locally with embedded_models=false (zero bundled neural weights and zero mandatory external embedding API calls), every store gateway keeps its memory index inside the local LAN perimeter while keeping RAM and storage overhead flat.
How Does LongMemOS Isolate Store Tenants on RK3588 ARM64?
LongMemOS isolates store tenants on ARM64 hardware by binding each store identifier to a dedicated local storage namespace and enforcing tenant-scoped MCP routing before any index read executes. Because each tenant namespace lives in its own isolated SQLite/local file boundary under an unprivileged Linux service user, a malformed prompt from one store terminal cannot traverse into another store's historical records.
Below is a minimal ARM64 systemd and multi-tenant configuration pattern for running a local LongMemOS MCP node on an RK3588 gateway (aarch64-unknown-linux-gnu):
# /etc/systemd/system/[email protected]
# Instantiate one isolated runtime per store: systemctl enable --now longmemos-edge@store_042
[Unit]
Description=LongMemOS Edge Memory Node (%i)
After=network-online.target
[Service]
Type=simple
User=longmemos
Group=longmemos
WorkingDirectory=/var/lib/longmemos/tenants/%i
ExecStart=/usr/local/bin/longmemos-server \
--tenant-id=%i \
--data-dir=/var/lib/longmemos/tenants/%i/store.db \
--embedded-models=false \
--replay-topk=100 \
--replay-max-topk=200 \
--bind=127.0.0.1:4100
Restart=on-failure
RestartSec=3s
ProtectSystem=strict
ReadWritePaths=/var/lib/longmemos/tenants/%i
PrivateTmp=true
NoNewPrivileges=true
[Install]
WantedBy=multi-user.target
Because ProtectSystem=strict and ReadWritePaths restrict each service instance to /var/lib/longmemos/tenants/%i at the Linux kernel level, tenant isolation does not rely on prompt instructions alone. Even on a compact RK3588 board sharing compute with local POS sync daemons or camera pipelines, the ~3.1 MB binary footprint leaves headroom for other edge workloads.
How Does O13 Evidence-Boundary Replay Prevent Numeric Hallucination?
O13 Evidence-Boundary Replay prevents numeric hallucination by pairing structured memory lookups with an immutable raw_replayable_candidate index (default topK=100, hard cap 200) that returns exact source file spans alongside every recalled fact. Traditional memory layers compress older conversations into rolling natural-language summaries; once a summary drops a decimal digit or a revision date, downstream reasoning cannot recover it.
In LongMemOS, every operational rule update records a temporal supersession pointer while keeping the underlying verbatim transcript line addressable. When an MCP client queries a store's shelf-life markdown or maintenance log, the server returns both the active state and the auditable source span:
{
"jsonrpc": "2.0",
"id": "req-store042-20261121-01",
"method": "tools/call",
"params": {
"name": "memory_recall_o13",
"arguments": {
"tenant_id": "store_042_dalian",
"query": "chilled snack lot LOT-20261114-B markdown trigger window and cold-case alarm threshold",
"topK": 100,
"include_raw_replayable_candidate": true
}
}
}
{
"jsonrpc": "2.0",
"id": "req-store042-20261121-01",
"result": {
"tenant_id": "store_042_dalian",
"candidate_pool_size": 100,
"hard_cap_enforced": 200,
"verified_facts": [
{
"entity": "LOT-20261114-B",
"attribute": "markdown_trigger_days_remaining",
"active_value": "14 days (30% off auto-tag)",
"superseded_record": "10 days (replaced 2026-11-05)",
"o13_source_replay_candidate": {
"document": "ops/store_042/expiry_policy_202611.md",
"line_range": "44-49",
"verbatim_span": "Batch LOT-20261114-B: effective 2026-11-05, trigger 30% markdown at 14 days remaining; cold-case high-temp alarm remains 6.5 C."
}
}
]
}
}
Because the model receives the verbatim o13_source_replay_candidate span rather than a paraphrased summary, auditing why an agent recommended a specific stock markdown takes one line-number check instead of re-reading weeks of chat logs.
What Did Our Internal Benchmark Show on Long-Horizon Recall?
In an internal test against the 500-question LongMemEval benchmark, LongMemOS recorded 463/500 (92.6%) accuracy using its local O13 evidence-boundary replay pipeline without bundled neural weights. Below is the structured six-element evidence summary documenting the exact scope and boundaries of our internal measurements:
| Evidence Dimension | Verified Internal Record |
|---|---|
| Measurement Target | LongMemOS standalone memory engine (MemoryBox 2.0 architecture) with O13 Evidence-Boundary Replay enabled |
| Environment & Target Arch | Customer-controlled Linux x86_64 servers, Windows 11 developer hosts, and RK3588 ARM64 edge gateways (embedded_models=false) |
| Sample Size / Baseline | 500-question LongMemEval benchmark suite (internal test evaluation) |
| Evaluation Method | Local candidate retrieval (raw_replayable_candidate / source_replay_candidate, default topK=100, hard cap 200) paired with temporal supersession tracking |
| Quantitative Result |
463/500 (92.6%) recall accuracy in internal test; ~3.1 MB binary distribution size with zero bundled model weights |
| Boundaries & Caveats | Internal test observation; end-to-end answer quality still depends on the upstream reasoning LLM configured by the operator and the cleanliness of ingested source logs |
How Can Engineering Teams Evaluate Edge Memory Without Vendor Lock-In?
Engineering teams can evaluate edge memory without vendor lock-in by running a single-tenant MCP pilot against anonymized operational logs before wiring live store or factory terminals. Start with three concrete checks:
- Verify exact-string retention: Feed 50 real SKU codes, serial numbers, or port revisions into your current RAG pipeline and compare verbatim recall against O13 span replay.
- Test temporal override behavior: Ingest a Q1 SOP followed by a Q4 amendment for the same equipment ID, then verify whether your agent cites only the active rule while flagging the prior value as superseded.
- Audit offline resilience: Disconnect outbound internet access on an ARM64 test node (such as an RK3588 board) and confirm local MCP memory reads continue without external embedding API timeouts.
Engineered by Dalian Fidwell Technology Co., Ltd., LongMemOS is designed to run entirely inside your own server room or edge hardware fleet. If your team wants architecture specifications or a scoped on-premise evaluation, submit your requirements directly through the LongMemOS Inquiry Form.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.