gptme 0.34.0: Memory, Safety, and More Models

gptme 0.34.0 is out. The headline is a proper memory system — not a bolt-on, but a redesigned, layered gptme.memory package that works the same way whether you’re running gptme, Claude Code, or Codex. Plus safety hooks that make headless autonomous agents easier to run responsibly.

Update, September 19: Do not use the v0.34.0 desktop assets for a fresh local setup. A clean launch of the released Linux AppImage found two blockers:

The Local server preset was rewritten to tauri://localhost instead of the bundled server. That fix includes the released-AppImage reproduction, a working manual-URL control, and four regression tests.

The frozen sidecar omitted dynamically imported hook modules. That fix records the released sidecar’s import failures and a clean startup from the patched PyInstaller build.

Both fixes merged on September 19, but v0.34.1 has not been published yet. The CLI package is unaffected by these two desktop defects.

Memory gets a real architecture

The old setup: memory was ad hoc. Some things landed in workspace context, some didn’t. Cross-harness recall was manual. There was no way to control what stayed always-on vs. what was selectively injected.

0.34.0 replaces this with gptme.memory — a proper layered store with a CLI:

# Save something across sessions
gptme-util memory save "project-context" "We use squash merges, master branch, conventional commits"

# Search it later
gptme-util memory search "merge strategy"

# Export for inspection
gptme-util memory export --format json

The layers matter: you can have always-on entries that inject on every session, keyword-triggered entries (same mechanism as lessons), and entries that respond to semantic search via gptme-rag. Each entry carries provenance — you can see what wrote it and when.

Cross-harness: this is the part that’s been missing. If you use gptme alongside Claude Code or Codex, memory entries now load across all three. The Codex/AGENTS.md integration means Codex picks up the same layered context gptme and CC already had.

Knowledge base migration: the old gptme knowledge CLI is deprecated. If you have existing knowledge JSONL, there’s a migration command:

gptme-util memory migrate-knowledge-jsonl your-knowledge.jsonl

Safety for headless agents

Two new features that matter for running autonomous sessions:

Trust-on-first-use for shell execution: when running in a project for the first time, gptme now gates shell execution behind an explicit trust confirmation. Subsequent runs in the same project skip the gate. This prevents a fresh clone from immediately running arbitrary scripts without your awareness.

TOOL_CONFIRM hook: a new hook type you can wire to a script that approves or denies tool calls before they execute. Useful for headless setups where you want to log or conditionally gate dangerous operations:

# gptme.toml
[hooks]
TOOL_CONFIRM = "scripts/confirm-tools.sh"

The confirm script receives the tool name and arguments, returns 0 to allow or non-zero to deny. Pairs with the existing ToolSpec.read_only flag — read-only tools auto-approve in the CLI confirm hook.

Agent infrastructure

A few changes that compound:

Mid-session tool switching: the harness now actually actuates request_tool_change — you can enable or disable tools while a session is running. Useful for tightening permissions mid-task.

Subagent persistence: the subagent registry persists across restarts and rehydrates on startup. Child conversations can be continued after a restart rather than starting fresh.

Context-scout pre-pass: an optional cheap-model pass identifies relevant files before loading them into context. Token savings for large codebases where the main model was spending budget on files that turned out irrelevant.

New models

The model list is updated:

deepseek/deepseek-v4.1-flash — fast and cheap
deepseek/deepseek-v4-flash-0731 — the July checkpoint
openai/gpt-6-astra — available via OpenRouter
glm-5.3-flash — GLM’s latest fast model
grok-4.6 — via xAI

Reasoning effort is now controllable for OpenAI-compatible providers:

gptme --reasoning-effort high "analyze this codebase"

The --track-tokens flag prints a running token count per LLM call — useful for diagnosing where context budget goes.

Developer UX

Auto-discover local providers: gptme now scans for local Ollama and LM Studio instances automatically. No config needed if they’re running on default ports.

Skills as slash commands: /skill:<name> now works in the TUI, same as it does in Claude Code. Consistent interface across runtimes.

Service init scaffold: gptme service init generates a systemd (Linux) or launchd (macOS) unit file for running gptme as a headless background agent. Takes care of the boilerplate.

Shell streaming: one-shot command output now streams while it runs instead of buffering. Visible progress for long-running commands.

New site

gptme.org got a design-v3 refresh, deployed as part of this release. Cleaner layout, better showcases what gptme actually does.

Get it

pipx upgrade gptme
# or
pip install --upgrade gptme

Full release notes at github.com/gptme/gptme/releases/tag/v0.34.0. Wait for v0.34.1 before installing the desktop app; the v0.34.0 release assets contain the first-run defects described above.

If you’re upgrading from a version before 0.33.x and using the gptme knowledge CLI, run the migration command before the old CLI is fully removed.