Two selected releases address different ways an AI agent can lose useful state: Claude-Mem returns a Codex subagent’s memory timeline after context compaction, while Mastra adds ownership claims and storage fences for durable-run recovery. The features solve separate problems. Both accounts come from their maintainers’ release notes; this newsroom has not installed either project.
Claude-Mem returns a subagent’s session observations after compaction
Claude-Mem captures observations from agent sessions for later retrieval. Its v13.39.0 release adds a hook sequence for Codex subagents. The release calls the returned material a memory timeline; the project’s pinned README describes session observations and chronological timeline views around those observations.
After context compaction, a PostCompact hook marks the thread using its session and agent IDs. At the thread’s next PostToolUse, the observation is recorded, the marker is consumed and the timeline is returned once as additionalContext. The release says Codex does not send SessionStart to subagent threads; that hook clears the marker so the root thread does not receive the same timeline twice. The behavior depends on Claude-Mem’s Codex integration and hooks being active; the newsroom did not test them. The repository code is Apache-2.0, but hosted observer, model-provider and data-handling terms are separate.
Mastra changes recovery safeguards and default model selection
Mastra is a TypeScript framework for AI applications and agents. Its v1.76.0 release addresses a recovery race in which one process can still be running a durable agent while another tries to recover that run. The release says the executing process owns a renewable run claim; the claim lapses about 30 seconds after that process dies. Recovery skips a live claim. A forced takeover stops the execution that held it.
Storage-level run fences can reject writes from an execution that has lost ownership. The workflow store records run ownership, while the memory store raises the matching fence. Full fencing requires both stores to support their part of this path. If only the workflow store supports fencing, messages, threads and working-memory writes can still land; if only the memory store supports it, the workflow store cannot provide the ownership claim a stale write must be checked against, so no writes are fenced. Unsupported custom stores default to no run-fencing support, and durable agents fall back to a PubSub lease.
The same release changes Observational Memory’s default model to auto, which may select a different provider or model based on available credentials. Teams that need fixed routing should pin and test the intended model. Most Mastra repository code is Apache-2.0, while ee/ directories have a separate Enterprise License.
What to check before adopting either change
For Claude-Mem, test repeated compaction in a root-and-subagent setup and confirm the right timeline returns once to the right thread. For Mastra, verify the actual production stores’ fencing support and simulate a stale process writing after a recovery takeover. These checks follow from the documented behavior; the newsroom has not run them. GitHub Picks is an editorial selection, not a GitHub ranking or independent product review.
Sources
- Claude-Mem repository; v13.39.0 release notes; repository documentation; repository license.
- Mastra repository; @mastra/core@1.76.0 release notes; run-ownership and storage-fencing implementation PR; repository documentation; repository license.
- Illustrative image: “Computer programmer typing code,” Free-Photos from Pixabay via Wikimedia Commons, CC0 1.0. It does not depict either project.
