日本語 Sign in Request an invite

Help

Tips for a memory that grows

How a memory grows depends on where it lives and how you ask. These are the small habits that help. None of them is required; pick what fits the way you work.

Reading time: about 7 minutesBefore you start: your AI client is connected

Kagura AI is the shared memory platform for people and AI agents.

On this page
  1. 1 Decide where memory lives
  2. 2 Move the notes you already have
  3. 3 Start and end of a session
  4. 4 How to teach it
  5. 5 How to ask
  6. 6 How much it recalls
  7. 7 Frequently asked questions

1Decide where memory lives

Most AI clients keep a memory of their own. You can teach both, or gather everything in Kagura. Both work; the choice decides how much sorting you do later.

Gather it in Kagura

Everything sits in one place, and the same memory is visible from another client, another machine, and your teammates. Switch the client's own memory off, or leave it unused. Coding tools switch theirs off with a single setting.

Split the roles and keep both

Leave personal preferences and phrasing to the client, and put work decisions, deadlines and know-how in Kagura. The memory in chat products can stay on. Teaching the same fact to both makes it unclear which one is current, so avoid only the overlap.

What to check, client by client

Claude (claude.ai and Claude Desktop)

Its own memory
“Generate memory from chats”, and “Search and reference chats” for past conversations.
To switch it off when gathering in Kagura
Settings → Your Claude, switch off “Generate memory from chats”. Choose Pause to keep what exists or Reset to delete it. For a single chat, open “+”, turn Memory off, then send.
Good to know
Reset cannot be undone. On Team and Enterprise plans, the organization's setting takes precedence.

Claude Code

Its own memory
An auto memory per project (~/.claude/projects/<project>/memory/).
To switch it off when gathering in Kagura
Turn off the toggle in /memory, or set CLAUDE_CODE_DISABLE_AUTO_MEMORY=1.
Good to know
CLAUDE.md is still read and the notes already written stay where they are. One toggle brings it back.

ChatGPT

Its own memory
Memory (saved memories, reference to chat history, content from connected apps).
To switch it off when gathering in Kagura
Settings → Personalization → Memory. Turning off “Reference saved memories” also stops chat-history reference. For a single conversation, use a Temporary chat.
Good to know
Turning off chat-history reference schedules the memory built from history for deletion within 30 days. Which controls appear varies by plan and region.

Codex (CLI, IDE extension, desktop)

Its own memory
Local memories under ~/.codex/. Off by default.
To switch it off when gathering in Kagura
Nothing to do if you left the default. If you turned it on, set memories = false under [features] in config.toml; /memories switches it per chat.
Good to know
If you keep both, memories.disable_on_external_context = true keeps chats that used MCP tools out of memory generation.

Cursor

Its own memory
The Memories feature was removed in version 2.1.17.
To switch it off when gathering in Kagura
Nothing to do.
Good to know
Leftovers can be exported with “Export memories” to an .mdc file and moved into Rules.

Hermes Agent

Its own memory
MEMORY.md and USER.md in ~/.hermes/memories/, loaded at session start.
To switch it off when gathering in Kagura
In config.yaml, set both memory.memory_enabled and memory.user_profile_enabled to false.
Good to know
External providers run alongside by default. With only one of the two set to false, the memory tool stays.

OpenClaw

Its own memory
MEMORY.md in the workspace, daily notes, and memory-core's search and dreaming.
To switch it off when gathering in Kagura
memory.search.enabled: false stops search; plugins.entries.memory-core.config.dreaming.enabled: false stops dreaming.
Good to know
openclaw memory reset keeps the files.

Menu names and settings keys follow each service's official documentation as of November 2026. If your screen differs, check that service's help.

2Move the notes you already have

Notes that built up in a client can move into Kagura as they are. Moving only what is still true, rather than everything at once, makes them easier to find later. Review the list before saving, and put them in the project's context.

Choose your client

Claude Code

The notes are in ~/.claude/projects/<project>/memory/. MEMORY.md is the index; each topic is its own file.

Read MEMORY.md and the topic files in ~/.claude/projects/<project>/memory/ and remember in Kagura Memory only what is still valid. One item at a time, with the point, the reason, and the date if known. List anything stale or duplicated for me to confirm before saving.

If it asks before reading outside the working directory, approve it. Replace <project> with the actual folder name.

Codex

The notes are in ~/.codex/memories_v2/ (~/.codex/memories/ on older setups). Memories are off by default, so there may be none.

Read the memory files in ~/.codex/memories_v2/ and remember in Kagura Memory only what is still valid. One item at a time, with the point, the reason, and the date if known. List anything stale or duplicated for me to confirm before saving.

If it asks before reading outside the working directory, approve it.

Hermes Agent

The notes are MEMORY.md (the agent's notes) and USER.md (your profile) in ~/.hermes/memories/.

Read MEMORY.md and USER.md in ~/.hermes/memories/ and remember in Kagura Memory only what is still valid. Keep USER.md as facts and preferences about me, and MEMORY.md as decisions and know-how. List anything stale or duplicated for me to confirm before saving.

If it asks before reading outside the working directory, approve it.

OpenClaw

The notes are MEMORY.md in the workspace (default ~/.openclaw/workspace) and the dated notes under memory/.

Read MEMORY.md in the workspace and the notes under memory/ and remember in Kagura Memory only what is still valid. One item at a time, with the point, the reason, and the date if known. From the dated notes, pick out decisions and facts only. List anything stale or duplicated for me to confirm before saving.

The dated notes add up; starting with the last few weeks keeps the review manageable.

Cursor

The Memories feature has been removed. Anything left can be exported from the command palette with “Export memories” to an .mdc file.

Read the exported memories.mdc and remember in Kagura Memory only what is still valid. One item at a time, with the point and the reason. List anything stale or duplicated for me to confirm before saving.

Put the exported file inside the open project so it can be read. Replace the file name with the real one.

Claude (claude.ai and desktop)

Claude shows a summary of what it remembers in Settings → Your Claude. Copy the entries you want to keep.

Below is what Claude's memory remembered about me. Remember in Kagura Memory only what is still correct. Keep preferences and facts separate, and pin the rules:

(paste here)

With many entries, batches of about 20 are easier to review.

ChatGPT

Copy the entries you want to keep from Settings → Personalization → Memory (Manage or Saved memories).

Below is what ChatGPT's memory remembered about me. Remember in Kagura Memory only what is still correct. Keep preferences and facts separate, and pin the rules:

(paste here)

With many entries, batches of about 20 are easier to review.

Other

Copy the entries you want to keep from the memory screen or files of the tool you use.

Below is what another assistant remembered about me. Remember in Kagura Memory only what is still correct. Keep preferences and facts separate, and pin the rules:

(paste here)

If the client can read files, you can start with “Read <path to the file>” instead.

After moving

List what you remember about this project. Tell me if anything is duplicated or out of date.

Review the result; anything you no longer need goes with “forget this”.

3Start and end of a session

Read back where things stood at the start, and leave what you decided at the end. That round trip is the shortest way to a memory that grows.

At the start

Pick up where we left off. Read the pinned rules, what is coming up, and what changed since last time.

Pins, upcoming dates and changes come back in one go.

At the end

Remember what we decided today, with the reasons. Add anything to carry over to next time.

The next session starts from here.

  • With the Claude Code plugin, the session opens with its default context announced, and /kagura-memory:session-start and /kagura-memory:session-summary do the same two things. The summary lists candidates and you choose what to keep.
  • Pin sparingly. Keep pins to the rules that must always apply, so what is read back each time stays small.

The Claude Code plugin on GitHub →

4How to teach it

  • Decisions with their reasons. “Remember that the billing service's database is PostgreSQL. The reason is…” lets the next session continue instead of deciding again.
  • Rules that must apply every time: pin them. “Remember as a rule for this project and load it at the start of every session: no deploys on Fridays.”
  • Anything with a date, with its date. “Remember that November 21 is the deadline” lets it surface at the right time when you ask what is coming up.
  • One context per project. Keeping work and personal, or project A and project B, apart keeps searches clean.

5How to ask

  • Ask for the reason. “Why did we decide that?” brings back the conclusion and, if they were saved, the options considered.
  • When something is not found, add a name or a keyword. “Search by keyword for …” tries a different way of looking.
  • Ask about change. “How much has it moved since last month?” reads numbers recorded as a series straight back.
  • If Kagura is not being used, add “in Kagura Memory” to the request.

6How much it recalls

How much it recalls, and what that costs, is on the benchmarks page, together with how it was measured.

See the benchmarks →

7Frequently asked questions

Is it safe to turn off Claude Code's auto memory?

Yes. CLAUDE.md is still read, the notes already written stay on disk, and the toggle in /memory brings it back. If Kagura is where your memory lives, switching it off keeps everything in one place.

Should I turn off memory in ChatGPT or claude.ai?

You don't have to. Their memory is a personal-preferences layer and works alongside Kagura. The settings for keeping one place are in the table above. Note that turning off ChatGPT's chat-history reference schedules that memory for deletion within 30 days, and Claude's Reset cannot be undone.

Can I move the notes I already have into Kagura?

Yes. A client that can read files can read and save them; for chat products, copy the entries and paste them. Move only what is still true, and review the list before saving. The prompts above are per client.

What happens if I leave both memories on?

It works, but the same fact ends up in two places and it becomes unclear which is current. Split the roles, or avoid teaching the same thing to both.

How much does Kagura recall?

Accuracy and cost against other approaches are on the benchmarks page, together with how they were measured.