Universal Memory Tool

Long-Memory Infrastructure For Codex GPT-5.5, AnythingLLM, And Local Agents

Universal Memory Tool turns huge chats, project folders, notes, logs, and documents into a private searchable memory workspace for AI assistants.

The app stores compressed local evidence blocks plus small searchable headers. Assistants search first, then hydrate only the proof they need instead of replaying an entire chat history.

Best tested path: Codex Desktop with GPT-5.5 or other Codex models through local STDIO MCP. This keeps the memory database on your machine.

Also useful for: AnythingLLM and other local assistants that can call local tools or MCP-style commands.

ChatGPT web note: ChatGPT app/connector mode is a separate remote-capable path. Use it only if your workspace supports it and you understand the privacy tradeoff.

Download Windows Alpha (.zip)

Version: alpha - Size: TBD - SHA256: checksum file - setup guide

Unsigned alpha build. Windows may show an Unknown Publisher or SmartScreen warning. Hashes and setup notes are included in the download.

What It Does

Stores large chats and project files as compressed local evidence.

Searches small headers first, then hydrates raw text only when needed.

Reduces repeated context pasting, token waste, and middle-context loss.

Keeps memory data on the user's machine; VPS is for license/auth only.

Tested with a 321M-character Codex chat ingestion flow.

Who It Is For

Codex users working with GPT-5.5 across long implementation sessions.

AnythingLLM and local assistant users who want a private memory backend.

Developers who need evidence-first recall instead of summary-only memory.

Researchers who need repeatable retrieval over large source material.

Teams testing MCP-based memory before building hosted infrastructure.

Use It With

Codex + GPT-5.5

The primary tested workflow. Install the app, open MCP Connect, copy the Codex form fields, add a custom STDIO MCP server in Codex, then restart Codex.

Use GPT-5.5 in Codex if available, but keep the same deterministic memory contract: policy first, raw evidence first, summaries after.

AnythingLLM

Use as a local memory backend when your AnythingLLM setup can call local tools or MCP-style commands. Keep the memory root on the same machine for private local mode.

Do not route private local memory through a public relay unless you intentionally accept that privacy tradeoff.

ChatGPT App Mode

Optional and separate from Codex local MCP. ChatGPT app/connector setups can be remote, so they need a deliberate relay/security design.

For private local memory, Codex Desktop and local assistants are the safer baseline.

Anti-Drift Workflow

Deterministic Long-Chat Flow

Start from the latest distilled summary and active policy.
Search narrow first, then hydrate only the raw evidence needed now.
Ingest full raw chats/files as compressed evidence before writing summaries.
Write ordered stage summaries, then one final summary with decisions, mistakes, and next actions.
Use latest/timeline records for current state before broad semantic search.

Why It Helps

Prevents summary-only drift.
Keeps exact commands/code recoverable.
Lets the assistant reason over the whole project without loading everything.
Makes mistakes and corrected decisions visible to the next chat.

Alpha Download

Download package includes installer EXE, required companion BIN, SETUP.md, release notes, public verify key, hashes, and MCP skill guidance.

Keep UniversalMemoryAppInstaller.exe and UniversalMemoryAppInstaller-1.bin in the same folder before running setup.

This alpha is unsigned to avoid code-signing cost before real user demand is proven.

Verify hashes from SHA256SUMS.txt or PACKAGE-SHA256SUMS.txt if Windows or your browser warns about the file.