HomeFeaturesApple appsBenchmarksReviewable memoryOpen sourceGitHub ↗Get MOOTx01 ↓
Memory tool · MOOTx01 v1.0.25 and later

A record you
can review

Anthropic now tells teams to build memory systems, and Claude can use a simple file-shaped interface to do it: one memory tool, one /memories folder, and familiar file operations.

MOOTx01 gives Claude that same interface while saving each memory as a record you can review — available in v1.0.25 and every release since.

The agent gets the simple memory handle. You get memory you can inspect, confirm, limit, withdraw, and carry forward.

Get MOOTx01 Read about MOOTx01 Memory Adapter
Compare the three memory integrations ↓
Available today Free for individuals Same memory_20250818 interface Every memory a reviewable record
Why This Matters Now

The role of long-term memory

Memory has existed in AI systems for a while. People have stored summaries, embeddings, preferences, transcripts, project notes, and many other forms of scratchpad state.

The model changed the timing. Fable 5 made memory more practical. Anthropic's guidance now treats memory as part of the agent loop, and the memory_20250818 tool gives Claude a simple way to use it across conversations.

A file-shaped interface is useful because everyone can understand it. A model can use files. A developer can inspect files. A product team can explain files.

But a file only tells you what text is present. Long-running agents need more than that.

The Missing Questions

Questions for a durable memory record

When an agent writes memory, the next session may treat that memory as working context. Before that happens, the user should be able to ask:

Six questions attached to one memory record A memory record in the centre, with six questions connected to it: who wrote this memory, was it confirmed by a person, what did it replace, what changed later, can it leave this machine, and should the next session trust it. ONE MEMORY RECORD Keep the project local-first; cite exact files. SOURCEmodel CONFIRMEDnot yet HISTORYrevision 2 EXPORTABLEno TRUSTreview first CONFIRM · EDIT · WITHDRAW · HISTORY 01Who wrote this memory? 02Was it confirmed by a person? 03What did it replace? 04What changed later? 05Can it leave this machine? 06Should the next session trust it?
01Who wrote this memory?
02Was it confirmed by a person?
03What did it replace?
04What changed later?
05Can it leave this machine?
06Should the next session trust it?
"MOOTx01 is built to keep those questions attached to the memory."
The Contract

The memory interface and its records

The same memory calls, stored as reviewable records Claude issues the six memory_20250818 operations: view, create, str_replace, insert, delete, rename. Each call travels into the MOOTx01 vault, where the memory is stored as a record with source, confirmation state, history, and export control, and a file-shaped response returns to Claude. CLAUDE · THE AGENT Keeps the interface Anthropic defined memory_20250818 · /memories THE SAME FILE-SHAPED CALLS FILE-LIKE RESPONSE RETURNS viewcreatestr_replaceinsertdeleterename MOOTX01 · THE VAULT Stores a record SOURCEmodel · ClaudeCONFIRMEDnot yetHISTORYrevision keptEXPORTABLEnoDELETEaudited, not erased
Anthropic memory interfacememory_20250818MOOTx01 storage vaultRecords · History · Audit · Withdrawal

Claude keeps the interface Anthropic defined

Claude uses the memory tool to work under /memories and receives the file-like responses it expects. Both the MCP memory-tool interface and the Python handler support this contract.

view create str_replace insert delete rename
memory.create /memories/project-notes.md
saved as reviewable record · source: model · unconfirmed

MOOTx01 changes what happens after the write

Each memory becomes a record the user can review. New model-written memories start unconfirmed. Edits keep older versions in the history. Deletes remove memory from active use while keeping a record of what happened.

Claude works through its familiar interface while users retain control of the records it creates.

After the Write

The lifecycle of a memory record

On create

Provenance

When the agent creates memory, MOOTx01 records that the model wrote it and keeps it unconfirmed until a person reviews it.

On edit

History

When the agent edits memory, MOOTx01 keeps the older version in the history. What it replaced stays visible.

On delete

Audit

When the agent deletes memory, MOOTx01 removes it from active use while preserving a record of the change.

On view

Familiarity

When the agent views memory, MOOTx01 shows Claude the file-shaped view it expects. Nothing about the interface changes.

"This gives agents a familiar way to work while giving users a practical way to manage trust."
Use It Today

Three ways to use reviewable memory

Choose the integration for your workflow: Claude Code’s project memory, a memory tool for your MCP client, or a Python handler for your own application. These are optional connections to MOOTx01; each serves a different purpose.

Claude Code · harness-memory

Bring your project memories into MOOTx01

Use MOOTx01 in place of Claude Code’s native project memory. Setup offers to import your existing project memory files, and new project memories are captured as records in your estate.

Enable and import project memories
mootx01 enable harness-memory

Requires a running MOOTx01 daemon. Setup asks before importing existing memories.

Disable and export project memories
mootx01 disable harness-memory

When you disable the integration, choose the export offer to return both imported memories and newly captured project memories to Claude Code’s native memory files.

MCP clients · memory-tool

Give your AI the familiar memory interface

Enable Claude-style memory operations for Claude and other MCP-capable clients connected to MOOTx01. Your AI can view, create, edit, delete, and rename memory under /memories, while MOOTx01 keeps the underlying records reviewable.

mootx01 enable memory-tool

Enable this on your installed MOOTx01 daemon and use it through a connected MCP client. Project-memory import and export are handled separately by harness-memory.

Explore the interface and records
Messages API · moot-memory

Add reviewable memory to your own agent loop

Install the moot-memory Python package to handle Anthropic’s memory_20250818 tool in your application. The drop-in handler preserves the tool contract and stores memory in your MOOTx01 estate.

pip install moot-memory

Connect the handler to a running MOOTx01 daemon, then wire it into your Messages API tool loop. Installing the package alone does not connect your application.

Python package & usage

MOOTx01 also connects through scripts, plugins, and server APIs. Explore all nine integration modes.

Available today in MOOTx01 v1.0.25 and later, and free for individuals to use.

ReleaseMOOTx01 v1.0.25 release notes (memory tool introduced)github.com → Packagemoot-memory on PyPIpypi.org → ReferenceAnthropic memory tool documentationplatform.claude.com → ReferenceAnthropic tool referenceplatform.claude.com → ReferenceAnthropic Fable 5 prompting guideplatform.claude.com →
Summarize this page with AI
Claude ChatGPT Perplexity Google AI