recall 0.4.2

recall's tools stop failing with database is locked. Search results and the chat index are unchanged.

Upgrading

Restart the agents that were open when you upgraded. Their recall MCP server still runs the release it started with, and keeps failing until the agent restarts. recall status lists them.

Benchmarks re-measured for 0.4.1 and 0.4.2. Search scores on every public benchmark are unchanged. Notes written by Claude Haiku 5.5 hold 90.6% of MemoryCode's conventions in their latest form, and loaded at session start they lift an agent on long histories from 32% to 75%. Agents take in about as many tokens per question as with 0.3.0 run the same day, plus the definitions of the tools 0.4.0 added. LongMemEval accuracy is 88.4%. On my own chats, gpt-6-luna's notes leaned ahead of Haiku's, within noise. Agent results are compared within a day, because the model's behaviour shifted between 12 and 16 October. docs/benchmarks/agent_tokens.py now fits one factor over any number of releases.

recall's tools no longer fail with database is locked. Since 0.4.0, on a machine with more than one Cursor build, every index refresh opened a write that nothing committed, and an agent's recall MCP server kept it open after its first tool call. Its later calls, and other agents' and hooks', waited 30 seconds for the index and failed. In the eval of questions about past work, a third of recall's tool calls failed this way. A refresh with nothing to write now writes nothing, a refresh commits what it wrote, and the chat and note tools close their connection to the index at the end of each call.