MemoryBox 2.0 · v2.0.4

Local shared memory for AI tools — trace past decisions back to the exact source

Save conversations, decisions and debugging notes from Claude Code, Cursor and Codex on your own computer or internal server. Ask from any connected tool and MemoryBox finds matching source passages before handing them to the model you already use—no built-in model, no rebuilding when you switch.

Hardware-verified · Win11 x86_64 · Linux x86_64 · ARM64 RK3588

UI illustration · sample data, not a real screenshot

Three steps: save → ask from another tool → check the source

Claude Code · Order service project
Raise the payment callback timeout from 500ms to 2s — the gateway often takes over 1s at peak.
Config updated, and the decision has been saved to MemoryBox.
Saved · scope: Order service project · 2026-10-14 09:12
What you do
You agree on a config change in Claude Code.
What MemoryBox does
Stores the original conversation with its time, source tool and project. It is searchable immediately.
What you get
A project memory with a traceable source.
Use cases

Who it is for

Teams whose key decisions and know-how are scattered across individual AI conversations — hard to find when needed, and hard to trust as current when found.

Engineering teams using several AI coding tools

The problem
Design trade-offs, API changes and fixes live in each person’s Claude Code or Cursor sessions. When the tool or the person changes, the questions start over.
How it is used
Each tool connects to the same project memory over MCP, and decisions made in sessions are saved automatically.
Result
New colleagues or a new AI assistant get the original wording, date and speaker when they ask.

Law firms, consultancies, healthcare and other confidentiality-bound organisations

The problem
They do not want case files, working papers or patient records sent to a third-party vector service just to make them searchable.
How it is used
Use the portable or on-premises server edition. Memories and attachments are stored encrypted on your own hardware.
Result
Search runs locally. If your AI tool uses a cloud model, only the retrieved passages are sent to it with the question.

International teams working in Chinese, Japanese and English

The problem
Specs are in Japanese, code comments in English and meeting notes in Chinese, so nobody can search across them.
How it is used
Built-in Chinese/Japanese/English tokenisation, Simplified/Traditional and kana normalisation, and unit conversion (0.5s = 500ms).
Result
A question in Chinese can find the parameter recorded in Japanese meeting notes.

AI PC and industrial gateway makers

The problem
Field devices have little memory and are often offline, so they cannot host database containers or models.
How it is used
Run the 3.55 MiB fully static server on ARM64 devices; idle memory use is about 2.25 MB.
Result
Devices can search past events and service records locally.
How it works

How MemoryBox works

MemoryBox stores and finds; the model you already use writes the answer.

  1. 01

    Save

    Your AI tool calls the write API during or at the end of a session. The original text is stored encrypted and is searchable right away. Screenshots, PDFs and drawings can be attached to the record.

  2. 02

    Find

    When you ask, MemoryBox searches from several angles at once — Chinese, Japanese and English keywords, related people and things, time, numbers and versions — merges duplicates and assembles the original passages in time order.

  3. 03

    Check and answer

    Passages carry their date, source and speaker role and go to the host model to answer. When old and new decisions conflict, the current one comes first and the history stays available. If nothing supports an answer, it says there is no matching record.

Test evidence

Figures you can check

These are test records for MemoryBox v2.0.4, with the category, conditions and limits of each. No third-party report has been published.

LongMemEval: 463 of 500 questions answered correctly (92.6%)
Internal test · public dataset
Test conditions:v2.0.4, no built-in model (embedded_models=false). Official scoring rules (±1 day allowed on time questions). Answering and grading used external models; the specific models, test date and number of runs are not stated in the source material.
Limits:Not independently verified. Results vary with the answering model. We can walk you through a re-run during commercial discussions.
For 457 of 470 answerable questions, the needed evidence was in the top 30 results (97.23%)
Internal test · offline retrieval
Test conditions:Same dataset; retrieval only, with no external model or network calls.
Limits:Measures whether evidence was found, not the final answer accuracy.
Hardware-verified on Win11 x86_64, Linux x86_64 and ARM64 RK3588
Internal hardware verification
Test conditions:Before the v2.0.4 release, 7 usage paths were run — portable drive operation, save and search, permissions, evaluation, legacy migration, cross-tool sharing, ARM64 and Win11 hardware — all passed.
Limits:macOS is still in validation; RK3576 and other ARM64 devices are assessed per project.
Release package ≈3.1 MB compressed; server 3.55–4.56 MiB unpacked
Measured package size
Test conditions:v2.0.4. ARM64 fully static server 3.55 MiB (3,724,384 bytes); Linux x86_64 server 4.56 MiB.
Limits:Excludes memory data and attachments.
About 4.70 GB for one million conversation turns
Structural estimate
Test conditions:Extrapolated from the measured text compression ratio and a typical turn length (about 270 turns a day for 10 years).
Limits:Excludes large attachments such as drawings and screen recordings; actual size depends on content.

The hero walkthrough and the examples in the technical reference use sample data. This table holds test records; the two are labelled separately.

Technical reference

Test breakdown and performance · How it compares with other approaches · The 8+2 search lanes · Architecture and background tidying · Access scopes and security

Test breakdown and performance

All figures are internal v2.0.4 test records and have not been independently verified.

LongMemEval evidence retrieval by question type (retrieval stage, not final answer accuracy)

Swipe horizontally to view table
Question typeQuestionsEvidence in top 30Notes
Knowledge updates (knowledge-update)7272 / 72Tells old and new states apart; current value first.
What the assistant said (single-session-assistant)5656 / 56Recovers parameters and steps the AI assistant gave.
Facts the user stated (single-session-user)6464 / 64Names, numbers and other facts found directly.
Reasoning across sessions (multi-session)121120 / 121Needs clues joined from several conversations.
Time reasoning (temporal-reasoning)127122 / 127Separates when something happened from when it was recorded.
Implicit preferences (single-session-preference)3023 / 30Offline result without host-supplied expansion terms; the weakest category.
Questions to decline (abstention)30—No evidence retrieval; counted in the 500-question total.

Performance and resource use

Internal measurements. The source material does not list the exact test hardware or data volume, so treat these as orders of magnitude.

Cold start (create store)
3.64 ms
Desktop / server. Reopening an existing store: 0.32 ms; RK3588: 2.38–3.20 ms.
Multi-lane search P99
29.35 ms
Write P99 over 200 consecutive writes: 25.93 ms.
RK3588 idle memory
2.25–2.98 MB
About 16.2 MB steady state on desktop. No GPU / NPU used.
Offline viewer
≈57.7 KB
Layout step for a 5,000-node graph: 3.85–5.18 ms. No network needed.

How it compares with other approaches

Each approach suits different conditions. Because the published tests used different models and setups, we do not compare scores here.

Swipe horizontally to view table

ApproachSuitsThings to weigh
Put the conversation history straight into a long model contextSmall amounts of material, one-off analysis, no long-term storage needed.The full history is resent with every question, so cost and wait time grow with length; the history does not move with you to another tool.
Cloud vector or graph memory servicesYou already use a given cloud, your data may live there, and you want managed operations.Usually tied to a specific embedding model; changing models can mean re-indexing. Your data is stored by the provider.
MemoryBox local retrievalData must stay local or on your network, several AI tools should share one memory, or it must run on offline devices.Answer quality depends on the connected model; existing vector clusters can be added in the enterprise edition; native macOS is still in validation.

The 8+2 search lanes

A “lane” is one angle for finding relevant records. Eight core lanes run at the same time, plus two optional lanes. Click a lane to see what it handles.

L1 · Chinese / Japanese / English keywords

All examples are sample data.

Tokenises all three languages, normalises Simplified/Traditional and kana, and splits camelCase and snake_case names in code.

Example question
What was the E_FSM_09 error in order_status_fsm.rs last week?
Record found
[dev session #412] fix(fsm): resolve E_FSM_09 stale lease transition

Rules for merging results

  • One slot per sourceSeveral fragments of one long conversation take a single slot, leaving room for other sources.
  • Original wording firstRecords backed by original text rank above inferred links.
  • Say when results are cutWhen matches exceed the return limit, the total is reported so the model does not treat a sample as the whole.

Architecture and background tidying

Written in Rust in five layers; each layer depends only on the ones below it.

  1. L5Integrations and viewerAdapters for six AI tools, ZeroOS event integration, offline viewer.
  2. L4ProtocolsREST v2 with a concurrency limit, 7 MCP tools, UAMP-v2 capability negotiation.
  3. L3Access scopesFive scope levels; writes, searches, tidying and handovers share one permission check.
  4. L2Search and tidyingIndex-on-write, 8+2 search, background tidying, original-text reader.
  5. L1Encrypted storageEncrypted containers, deduplicated attachment pool, accidental-format protection on the portable drive.

Background tidying (Dream)

Original records are never edited after they are written. Tidying only adds summaries and links, logged as batches.

  • Light tidying when idleMerges different names for the same thing (emails, ticket numbers, paths) and marks old/new replacements.
  • Periodic deep tidyingWrites period summaries, cross-language keywords and skill cards; the host model can help.
  • Always tied to sourceResults without a real source passage are not stored. Each batch can be rolled back without touching original records.

Original-text reader

  1. 1Keep the contextKeeps a few turns either side of a match instead of a single cut-out sentence.
  2. 2Put it in time orderMerges passages from the same source in the order they happened, with date and speaker role.
  3. 3Evidence firstAsks the model to cite text and time before concluding, and to say it does not know when there is no evidence. The core code contains no regex rules or question-ID branches aimed at test questions.

Access scopes and security

One set of permission rules covers writes, searches, background tidying and handovers. Revoked access takes effect without a restart.

Five access levels

  1. L1Company — Policies and shared standards for everyone
  2. L2Department — Engineering, legal, finance and so on kept apart
  3. L3Project — Cross-department project files and client correspondence
  4. L4Role — Know-how for architecture, operations, approvals and other roles
  5. L5Individual / single AI assistant — Personal sessions and drafts, visible only to the owner by default

Security design

Encrypted storage
Implemented

Records and attachments are compressed and stored encrypted; attachments are deduplicated by content.

Startup protection and key wiping
Implemented

Keys are wiped from memory after use. If a key file is missing or fails verification, the server refuses to start rather than overwrite existing data with a new key.

Accidental-format protection (portable drive)
First three layers implemented

Separate tool and data areas, an exclusive file lock while running, and a background guard process (< 5 MiB). The hardware write-protect interface is reserved for future hardware.

Hardware secure element (SE)
Interface reserved · planned

Keys support multiple slots, so moving to SE-equipped hardware later will not require re-encrypting existing data.

Editions and delivery

Three editions, chosen by team size and hardware

All three editions share one data format, so a personal memory can later be merged into an enterprise server. Pricing is quoted by edition, seats and implementation scope; we do not publish list prices.

Edition 1

Portable drive edition

256GB NVMe portable drive; software ≈3.1 MB compressed
For
Individuals who need their material with them: architects, developers, lawyers, consultants, researchers.
Runs on
Win11 x86_64 and Linux x86_64 computers (macOS in validation).
Main deliverables
  • 256GB NVMe portable drive with v2.0.4 preinstalled
  • Configuration templates for six AI tools
  • Offline viewer and user guide
Pricing basis
One-off hardware purchase, no monthly subscription; model costs are separate.
Edition 2

Edge device edition

3.55 MiB fully static server; ≈2.25 MB idle memory
For
AI PC and industrial gateway makers, in-store terminals, offline workstations.
Runs on
ARM64 Linux (RK3588 hardware-verified; other chips assessed per project).
Main deliverables
  • Fully static server binary
  • MCP and REST v2 interface documentation
  • Device integration guidance
Pricing basis
OEM preinstall licence or joint delivery, discussed by shipment volume and integration scope.
Edition 3

Enterprise server edition

4.56 MiB server
For
Engineering centres and finance, manufacturing and other enterprises where several departments and AI tools work together.
Runs on
Linux x86_64 servers or VMs; no GPU required.
Main deliverables
  • Server plus five-level access scope setup
  • Read-only migration tool for legacy data (verified on a 5.93 GB legacy store)
  • Optional ZeroOS integration and enterprise vector add-on
  • Operations manual
Pricing basis
On-premises licence, with optional annual maintenance.

Supported platforms

Release package ≈3.1 MB compressed. Unpacked, the server is 3.55–4.56 MiB depending on platform, and the offline viewer is ≈57.7 KB.

  • Windows 11 x86_64Hardware-verified

    Runs as a single file with no extra DLLs; exclusive lock active while running.

  • Linux x86_64Hardware-verified

    4.56 MiB server; out-of-scope blocking and read-only legacy migration verified.

  • ARM64 RK3588Hardware-verified

    Fully static 3.55 MiB; service ready in 6.6–17.4 ms.

  • macOSIn validation

    Not yet a delivery commitment. Mac users can connect through a Linux node on the LAN.

What the quote is based on

MemoryBox has no built-in model and charges no embedding API fees. A deployment involves these parts:

MemoryBox licence or portable drive
Quoted by edition, seats or shipment volume.
Model costs
Use your existing model subscription or local compute, paid directly to the model provider.
Optional enterprise vector service
Enterprise edition only; not needed for core features.
Annual maintenance
Optional for the enterprise edition: version updates, access-scope planning advice and incident response.

Pilot evaluation steps

  1. 1
    Leave your details

    Tell us which hardware and AI tools you use; everything else can wait for the call.

  2. 2
    Confirm platform and integration

    An engineer confirms the platform, integration protocol and how access scopes should be split.

  3. 3
    Demo with your samples

    We show saving, finding, checking sources and out-of-scope refusal on your sample material, and can re-run LongMemEval with you.

  4. 4
    Agree edition and quote

    Based on the pilot, we confirm the edition, deliverables, maintenance terms and quote.

FAQ

Frequently asked questions

With no built-in model, who writes the answers?

The model behind your connected AI tool — Claude, GPT or a local open-source model, for example. MemoryBox finds the relevant original passages and puts them in time order; the model answers from them. That is why switching models does not require rebuilding your memory, but answer quality depends on the model you use.

Does data leave my computer or network?

Memories and attachments are stored on your computer, portable drive or internal server, search runs locally, and no external embedding API is called. If your AI tool uses a cloud model, the passages retrieved for a question are sent to that model with it. For fully offline use, connect a local model.

Can I start with the portable edition and move to an enterprise server later?

Yes. All editions share one data format and permission model, so a personal memory can be merged into a department or project scope on the enterprise server. Legacy stores can be migrated read-only and incrementally; in internal testing a 5.93 GB legacy store was migrated without changing the source files.

Can background tidying corrupt the original records?

Original records are not changed. Tidying only adds summaries and links, each of which must point to real source text or it is not stored, and every batch can be rolled back.

Does it run on a Mac?

A native macOS build is still in validation and is not yet a delivery commitment. For now, Mac users can connect to a Linux node on the LAN over MCP or REST v2.

Next step

Book a demo or pilot evaluation

Only your name and a contact detail are required; everything else is optional. You can also call or email us.

What happens after you submit

  1. 1Your inquiry is saved on our server.
  2. 2Our team reviews it during business hours and contacts you the way you asked.
  3. 3For anything urgent, call +86-411-81981709.
Dalian Fidwell Technology Co., Ltd.

MemoryBox inquiry

* required

We use your details only to respond to this inquiry and keep them for up to 24 months. You can ask us to show or delete them at any time. Full privacy notice