​

Current limitations

• Only Claude Code, Codex, and OpenCode have capture integrations.
• OpenCode can be used as a non-interactive memory-processing provider but does not receive automatic hook context.
• Agent event coverage differs. Termyte cannot store fields an agent never sends.
• Redaction is based on known keys and patterns. It cannot guarantee that every secret is removed.
• The Viewer exposes task records through read-only endpoints but has no task-management screens.
• The public task CLI does not expose every stored task record type. Requirements, decisions, and failures are available through the service but lack dedicated CLI commands.
• Import, export, bulk deletion, hosted synchronization, and Claude-Mem migration are not implemented.
• Execution projections cover captured prompts, completed tools, commands, and file reads or modifications. They do not reconstruct every possible tool lifecycle or file operation.
• Checkpoints describe Git state and changed paths; they do not store patch contents.
• Historical memories can become stale. Current code and current task state remain authoritative.
• File freshness checks use path existence and current Git changes. They do not prove that every claim in an old memory is still correct.
• Context attribution is conservative evidence, not proof that Termyte caused a task to succeed or fail.
• Passing tests and packed-install recovery prove software behavior, not real-world productivity gains.