Memory & Skills

Write it down once, and every agent reads it

Your project’s rules, conventions and pitfalls don’t need a copy in CLAUDE.md, AGENTS.md and GEMINI.md. Pyloxa keeps them on your computer and hands them to every agent it dispatches. Anything new an agent learns goes to Pending first, and counts only after you accept it.

One memory, read by all three CLIs

Click “Remember” in Shared memory, or save something from a conversation. Each memory has a scope (global or this project) and a kind: rule, preference, decision, fact or pitfall.

  • When Pyloxa dispatches a task, plans with the brain, or runs brain review or cross-testing, it adds the relevant memories to the prompt, labeled as background rather than instructions.
  • The order is pinned memories, then rules and preferences (project before global), then up to 12 of the most relevant facts, decisions and pitfalls.
  • Each dispatch adds up to 3,000 characters by default; you can choose 1,000 or 6,000. Each memory is at most 500 characters.
  • The run record says how many memories were included.
Shared memoryThis projectRemember
RuleAPI responses always use camelCasePinned
PreferenceTests live in tests/ and use Vitest
PitfallIntegration tests on Windows need --runInBand
3 shared memories were included in this run

Agents suggest, you approve

While they work, agents can search memory and suggest new memories through tools Pyloxa provides. Every suggestion goes to Pending first, and no agent sees it until you accept it.

  • Each task can suggest at most 5 memories, and the brain adopts at most 3 per turn, so memory doesn’t balloon.
  • Text that looks like a key or token is rejected and never stored.
  • To save time, you can let a project “Auto-accept facts and pitfalls agents propose”. It’s off by default, and rules and preferences always need your approval.
  • The memory tools are served only on 127.0.0.1, with a token for each run.
Pending (2)
From Codex · Fact
Points are calculated in src/points/; don’t touch legacy/
From Claude Code · Pitfall
Start the mock server before the E2E tests
AcceptEdit and acceptDiscard

Also there when you use a CLI in your terminal

Memory can come along when you open a CLI directly, without Pyloxa. “Sync to CLIs” writes the project’s memory as a Pyloxa block in AGENTS.md and adds @AGENTS.md to CLAUDE.md; Gemini CLI can be set to read the same file.

  • You see the diff first, and nothing is written until you click. Pyloxa never commits for you.
  • “Also write to each CLI’s user-level memory” is off by default. When you turn it on, Pyloxa backs up ~/.claude/CLAUDE.md, ~/.codex/AGENTS.md and ~/.gemini/GEMINI.md before the first write, and removes its block when you turn it off.
  • It works the other way too: import from the CLIs’ own memory files. Pyloxa reads only a fixed list, such as the project’s CLAUDE.md, .claude/rules/*.md, AGENTS.md and GEMINI.md, and Codex Memories. It never reads credentials, settings or chat history.
  • Memory can be exported as Markdown or JSON and imported from a file.
Sync to CLIsPreview
AGENTS.md
+ <!-- pyloxa:memory start -->
+ - API responses always use camelCase
+ - Tests live in tests/ and use Vitest
+ <!-- pyloxa:memory end -->
CLAUDE.md
+ @AGENTS.md
Nothing is committed for youWrite

Skills: see what each CLI can actually read

The Skills inventory lists the skills each CLI really sees, in a skill-by-agent table that shows what’s missing and what’s duplicated. That part is free.

  • Skills libraryPro: keep your skills in one place, ~/.agents/skills, and sync additions, edits and removals to each CLI.
  • AssignmentPro: choose which skills go to which agent, per project or per profile.
  • Dispatch packsPro: bundle rules, command suggestions, conventions, your workflows and skills into one file to share. Importing previews each item, your existing rules are never overwritten, and a workflow with the same name keeps yours unless you choose otherwise.
  • Gemini CLI runs don’t see user-level skills; they get only the skills assigned to them through Pro.
Skills inventoryMockup
SkillClaude CodeCodexGemini CLI
api-styleSees itSees it—
db-migrationSees it——
release-notesTwo versionsSees it—
ui-review—Sees it—

Free and Pro

Shared memory is entirely free. Pro adds managing and sharing skills in one place.

FreePro
Shared memoryRemember, Pending and approval, memory in every dispatch, sync to CLIs, import from CLI memory files, export and import; global shared memory also travels with your exported settingsSame as free
SkillsSkills inventory and the skill-by-agent tableA skills library synced to every CLI, assigned per project or profile
SharingProject rules can live in the repo’s .pyloxa/ folderDispatch packs: rules, command suggestions, conventions, your workflows and skills exported and imported together

Questions

Where is shared memory stored? Does it go to the cloud?

On your computer. It’s never sent to us and isn’t part of the anonymous usage statistics. The only time it leaves is when relevant memories go with a task’s prompt to wherever that CLI already sends it.

Can an agent write something wrong into memory?

Not in a way that takes effect on its own. Agents’ suggestions go to Pending, and no agent sees them until you accept them. Only if you turn on “Auto-accept facts and pitfalls agents propose” for a project are facts and pitfalls accepted automatically; rules and preferences always need your approval.

I already have CLAUDE.md or AGENTS.md. Do I still need this?

You can import them. Pyloxa reads a fixed list of memory files, and from then on one memory serves all three CLIs. When you also want it in the terminal, “Sync to CLIs” writes it back into a Pyloxa block in AGENTS.md, after showing you the diff.

Is the Skills inventory free?

Yes. Seeing which skills each CLI can read is free. The skills library, assignment per project or profile, and dispatch packs are Pro.

Try it on your own project

Pyloxa is a free Windows app. It runs the Claude Code and Codex CLIs you already use, on your own subscriptions or API keys.