agnostic-ai v0.81.0: one memory every AI tool shares

Agnostic-ai released v0.81.0, adding a shared memory system that loads project memory from `.agnostic-ai/memory/` and personal memory from `.agnostic-ai/local/memory/` into Claude Code, Codex, Copilot, Cursor, Gemini CLI, Qoder, Factory, OpenCode and Kilo Code at session start. The release also makes `lint` and `sync` fail on MCP `env` or `headers` values that are neither `${VAR}` references nor marked `!literal`, and stops sync on Claude agent names starting with `-` or containing `:`, which Claude Code skips. Upgrading requires running `agnostic-ai migrate --only secrets` and `agnostic-ai upgrade --requires`, and users enabling memory must upgrade the `agnostic-ai` binary on their PATH because the session-start hook runs that copy.

Upgrading from v0.80.0 Install it the way you installed the previous one, then fix the two new checks before you sync. 1. MCP plain values. lint and sync stop on an MCP env or headers value that is neither a ${VAR} reference nor marked literal . Run agnostic-ai migrate --only secrets , then fix the values it leaves to you secrets and plain settings https://agnostic-ai.org/docs/spec-format/mcps/ plain-settings . 2. Claude agent names. Sync stops on an agent whose name starts with - or contains : , because Claude Code skips that agent. Rename it, or set a valid x-claude.name . Then run this from the project root: agnostic-ai upgrade --requires It pins requires and the schema URL to 0.81.0 and syncs. If you keep a range or minimum in requires , keep it and run agnostic-ai sync instead. If you turn on memory, also upgrade the agnostic-ai on your PATH, not only a copy pinned in the project. The session-start hook runs the one on PATH, and an older one loads no memory. What agnostic-ai changed One memory for every tool Add the memory built-in and sync once: requires: " =0.81.0" builtins: handoff, memory Each tool now loads two short lists of facts when a session starts: project memory in .agnostic-ai/memory/ , committed for the team, and personal memory in .agnostic-ai/local/memory/ , kept out of Git. A rule tells each tool to save your preferences and corrections to personal memory, and to propose team facts before it saves them. Claude Code's own memory saves into the same folder. If Codex learns that CI runs only on Ubuntu, Claude Code and Cursor know it next session. Claude Code imports the lists, OpenCode and Kilo Code list them in their config, Codex, Copilot, Cursor, Gemini CLI, Qoder, and Factory load them with a hook, and every other tool reads them through the rule. See Shared memory https://agnostic-ai.org/docs/memory/ . One personal memory across worktrees By default each checkout keeps its own personal memory, and removing a worktree deletes what it saved. Set this in agnostic-ai.local.yaml to share one folder across every worktree of the repository: memory: personal: repo Sync gives Codex, Gemini CLI, Qoder, and the Cursor CLI write access to that folder, and works when ~/.agnostic-ai is a link to another folder. Tools that only get the rule find it with agnostic-ai memory path : bash $ agnostic-ai memory path project ~/code/app/.agnostic-ai/memory personal ~/.agnostic-ai/local/memory/app-fbcd84e8 See One store per repository https://agnostic-ai.org/docs/memory/ one-store-per-repository for which tools need the managed .gitignore block. Check and repair memory agnostic-ai memory lint , index , and list check, rebuild, and list memory. lint and doctor flag a list over 100 lines, a broken link, and text that looks like a secret. Fixes - Sync keeps hooks you wrote by hand on Claude Code, Codex, Cursor, Gemini CLI, Qoder, and Factory when it adds its own. - .cursor/cli.json now always has both allow and deny . The Cursor CLI refused a file with only one, so its rules never loaded. - A Copilot hook with a Windows command runs on Windows. - lint counts the AGENTS.md that CLAUDE.md imports in Claude Code's word total. Full list: v0.81.0 release notes https://github.com/Chemaclass/agnostic-ai/releases/tag/v0.81.0 . What changed in the targets Claude Code skips agents with some names. Its subagents page https://code.claude.com/docs/en/sub-agents checked 2026-10-07 says a name that starts with - , contains : , or is longer than 256 characters makes Claude Code skip the file and log an error. Claude Code 2.1.292 added the length limit. This release stops sync on such a name instead of writing an agent that never loads. Kilo Code ignores project-listed files outside the project. In Kilo v7.8.8, instructions entries from a project kilo.jsonc cannot read files outside the project root, and a blocked file is dropped without a warning source https://github.com/Kilo-Org/kilocode/blob/v7.8.8/packages/opencode/src/session/instruction.ts . With repo mode, sync no longer lists the personal folder there; Kilo finds it through the rule, or you add it to ~/.config/kilo/kilo.jsonc . Kilo v7.8.7 also added a memory model setting, which a settings spec can already set through x-kilo . Antigravity retires current workflows on October 19, 2026. Its workflows page https://antigravity.google/docs/ide/workflows checked 2026-10-07 says current workflows stop working on that date. agnostic-ai emits no Antigravity workflows, so synced output does not change; move any workflow you wrote by hand before then.