No Embedded Model Weights
embedded_models=false keeps the binary ~3.1 MB and decoupled from GPU VRAM. Upgrade your LLM anytime with zero database migration.
Save conversations, codebase conventions, and debugging notes from Claude Code, Cursor, Codex, and ZeroOS workers on your own workstation or intranet server. Query from any connected client and LongMemOS retrieves original timestamped source passages before your chosen LLM responds — no built-in model weights, no re-indexing when switching models.
Brand update notice: MemoryBox 2.0 operates globally as LongMemOS (LongMemOS.com). Internal LongMemEval figures below reflect Fidwell testing on v2.0.4 without third-party audit.
Illustrative sample data showing how cross-tool recall avoids vague summaries.
Work inside Claude Code or Cursor on authentication fixes, schema constraints, or deployment incidents.
Ingests structured event records with timestamps, workspace tags, and lexical/semantic hybrid indexes locally.
Original decision wording stays intact in your local store — never overwritten by a lossy model summary.
memory_write({ scope: "repo/auth", decision: "Rotate JWKS keys every 24h; keep kid v1 during drain" })
Switch to Codex, Cursor, or a ZeroOS background worker two weeks later and ask why an old key stays valid.
Runs multilingual query expansion and recency-aware retrieval across stored events via MCP tool call.
Works identically across MCP-compatible coding clients and local HTTP integrations without vendor lock-in.
mcp.call("memory_search", { query: "why does kid v1 stay active during rotation" })
Ranked original snippets with record IDs, dates, and surrounding lines ready for inspection.
Pass retrieved evidence to GPT, Gemini, Claude, or local models — swap models next month without migrating memory.
Auditable grounding eliminates invented project history when teams hand work across people or agents.
[hit #1 | 2026-11-02] "keep kid v1 during 24h drain to avoid invalidating active mobile tokens"
Separating durable retrieval from reasoning models keeps your memory fast, small, and portable.
embedded_models=false keeps the binary ~3.1 MB and decoupled from GPU VRAM. Upgrade your LLM anytime with zero database migration.
Plug into Claude Code, Cursor, Codex, Hermes, and custom Python/TypeScript workflows through documented MCP endpoints.
Search English codebase discussions using Japanese or Chinese phrasing — multilingual teams share one unified memory ledger.
Every memory hit returns source document identifiers, timestamps, and verbatim passages so engineers verify facts immediately.
Keep proprietary code rationale, client constraints, and incident post-mortems strictly inside your office LAN or Linux server.
Powers long-running multi-agent jobs inside ZeroOSAI.com so planners and executors reuse validated facts across runs.
LongMemOS solves persistent context across tools. When you also want an AI coordinator that splits multi-step tasks, assigns Linux executors, tracks state in PostgreSQL, and retries failed steps automatically, pair LongMemOS with ZeroOS AI.
Transparent reporting of measured dataset scores and physically verified hardware targets.
| Subject | Version / Scope | Recorded Outcome | Conditions & Boundaries |
|---|---|---|---|
| LongMemEval Public Benchmark (500 Qs) | MemoryBox / LongMemOS v2.0.4 | 463 / 500 (92.6%) | Internal automated test run with embedded_models=false and configured external evaluator LLM; not independently verified by a third party. |
| Desktop & Server Binary Builds | v2.0.4 release archive | Win11 x86_64 · Linux x86_64 (~3.1 MB) | Physically verified locally and on Linux servers (3.55–4.56 MiB unpacked on disk). |
| Edge ARM64 Appliance | ARM64 RK3588 Linux | Hardware-verified | Tested on RK3588 industrial/AI box hardware for LAN-shared team memory; macOS desktop build currently under validation. |
Deploy as lightweight workstation software, intranet container, or ready-to-plug hardware.
Single lightweight binary + local data store for developers using Claude Code, Cursor, and Codex daily.
Multi-workspace deployment inside company servers so engineering, QA, and support agents query unified memory.
Pre-installed quiet LAN appliance built around verified RK3588 hardware — plug power and Ethernet in your office.
MemoryBox 2.0 is the hardware/appliance lineage in China. Globally, we use LongMemOS (LongMemOS.com) to clearly signify a dedicated Long-Term Memory OS for AI agents and avoid naming confusion with unrelated consumer photo-keepsake websites.
No. LongMemOS runs locally on your PC, office LAN appliance, or company Linux server. Your stored memory remains on storage you control.
Your stored history, indexes, and MCP interface stay untouched because LongMemOS stores structured records independently of any LLM vendor.
Win11 x86_64, Linux x86_64, and ARM64 RK3588 are verified. macOS support is currently undergoing validation — contact us to join validation testing.
Share how your team runs coding tools or internal automation today. Only your name and one contact method are required.
All consultation records go directly to Fidwell engineering support. We respond during standard business hours (UTC+8).
Email: fidwellservice@fidwell.cn
Tel: +86-411-81981709
Address: Room 1303, Bank of East Asia Tower, No. 7 Renmin Road, Zhongshan District, Dalian, Liaoning, China