Intel / repo

crewAI 1.15.29

Source: github.com
Author: joaomdmoura
Date: Published Oct 23, 2026

Context

crewAI is an open-source Python framework for building systems in which several AI agents work together (GitHub). Agents in crewAI can store memories and recall them later, and a memory can be marked private, meaning only the source that created it should see it. Release 1.15.29 (release notes) fixes a bug that made the private flag ineffective for memories saved without a source.

The recall code showed a private memory to a caller whenever the memory's source matched the caller's source. A private memory saved without any source had an empty value there, and a caller that did not name a source also had an empty value, so the two counted as a match. Every default recall therefore received memories marked private, as long as they had been saved without a source. The fix changes both recall paths, the quick lookup and the deeper search, so a private memory is only shown when the caller names a real source that matches; the new test uses a stored API key as the private record that must stay hidden (commit).

Anyone who saved private memories without a source should know that recall now hides them from ordinary calls, which is the intended behavior. Code that needs them must name the matching source or ask for private records explicitly.

Key takeaways · AI-distilled

The crewAI command can now run the crew itself to reach a verdict in one command, keeps the evaluation's full data in the project, and can export evaluations as files.
Tool actions now honor their failure policies and get a fresh usage budget on each kickoff, and a separate fix stops agents from repeating the same tool call.
was refactored into a BaseLLM capability, and LanceDB storage fixes escape scope and id values in filters and correct the order in which memory records are retrieved.
The release adds Azure DocumentDB vector search, a read-only notebook text loader and a standard-library JSONL loader, and the CLI can now search platform applications and select an organization during login.

Terms in this piece · Glossary

eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.

Read the source github.com