LLM 0.32a1 Fixes SQLite Tool-Calling Bug in 2026: Restore AI Agent Memory Now

LLM 0.32a1, the latest update to Simon Willison’s open-source AI CLI tool, resolves a critical SQLite bug that broke multi-turn tool-calling sessions. Released in April 2026, this fix ensures AI agents can now reliably preserve conversation state—even after interruptions—making it indispensable for developers automating research, coding, and data workflows.

Why This Bug Broke AI Agent Workflows

Before version 0.32a1, LLM’s SQLite-based session storage failed to deserialize complex tool-calling chains. When users paused a workflow—like querying a database, fetching files, then running a shell command—the agent lost context upon restart. This resulted in incomplete outputs, repeated tool calls, or silent failures.

For AI engineers building automated agents, this wasn’t just an inconvenience—it eroded trust. Imagine an agent generating weekly reports by chaining SQL queries, API calls, and document exports. One broken session meant hours of lost work.

How the Fix Works: SQLite Persistence Enhanced

The 0.32a1 patch overhauls how LLM serializes tool parameters and conversation history into SQLite. It now uses JSON functions (available since SQLite 3.53.0) to cleanly encode function calls, arguments, and return values as structured JSON blobs, avoiding string truncation and encoding errors.

Simon Willison confirmed in GitHub issue #1426: "We now validate the entire state tree before committing to disk. If any part fails to serialize, the session is rolled back—no more corrupted states."

Real-World Impact: CLI Automation Just Got Reliable

With this fix, LLM becomes a production-ready tool for:

Automated data curation: Pulling, cleaning, and visualizing datasets across tools
Code assistants: Maintaining context across multiple file edits and test runs
Research agents: Chaining PubMed queries, PDF extraction, and summary generation

Users report up to 70% fewer failed agent runs after upgrading to 0.32a1. For teams relying on LLM as a background automation layer, this is a game-changer.

Why Open-Source AI Tooling Needs This Update

LLM’s philosophy—incremental, user-driven, lightweight—makes it a favorite among developers who distrust bloated AI platforms. Unlike cloud-only agents, LLM runs locally, uses SQLite for state, and integrates with Ollama, OpenAI, and Anthropic models.

With this fix, LLM joins the elite tier of AI tooling that treats agent memory as core infrastructure—not an afterthought. It’s no longer just a CLI; it’s a persistent, state-aware AI assistant.

As AI agents become standard in developer toolchains, reliable session persistence will be non-negotiable. LLM 0.32a1 sets a new benchmark for open-source AI tooling in 2026.