v2.22.0

Added

• ASCII-art favicon — The dashboard browser tab icon is now a stacked [---] / [M0] / [---] mark in the design system's cream/mono style (SVG, crisp at any size). (Entry moved from the 2.20.0 section, where it was listed before the feature actually shipped.)
• Independent switches for the learning pipeline — new promptTrackingEnabled and profileLearningEnabled config flags (both default true). Prompt persistence is no longer coupled to chatMessage.enabled, so prompt tracking can be off while memory injection stays on, and automatic profile learning on session idle can be disabled on its own. Together with the existing injectProfile and autoCaptureEnabled, every stage — prompt persistence, profile learning, profile injection, auto-capture — is now independently switchable while the Web UI, explicit memory tools, and semantic search keep working. (Requested by @ovizii in #55.)

Changed

• Dashboard clarifications — the memories stat card reads "global · project" instead of "user · project" (user-scope means cross-project memories, unrelated to the profile feature); the Memories filter chip reads "memories + prompts" instead of "all"; clicking a memory row's [+]/[?] marker or text now opens the detail view, in both the Memories list and search results (the marker looked like an expand button); the pin button tooltip explains pinning protects a memory from cleanup and ranks it first. (Raised by @ovizii in #55.)

Fixed

• User profile identity is now stable without an email — previously each learning run minted a fresh random anonymous identity, so one user accumulated a new profile per batch (four profiles from forty prompts in one report). All profile lookups now share one resolver: the git email when available, otherwise the most recent active profile (existing duplicate profiles self-heal onto it), otherwise a single stable anonymous identity. Saving an explicit preference from the profile tool also no longer fails when no email can be resolved. (Reported by @ovizii in #57.)
• Profile learning drains its backlog — previously only one batch (default 10 prompts) was analyzed per idle session, so a 1500-prompt backlog took roughly 150 idles. Learning now keeps analyzing FIFO batches until fewer than userProfileAnalysisInterval prompts remain or userProfileMaxBatchesPerIdle batches (default 5) are done, with a single summary toast per run. (Reported by @ovizii in #57.)
• Profile language stays pinned — a mixed-language batch could flip the profile's language between runs. Updates now keep writing in the existing profile's language; new profiles are created in the dominant language detected across the batch. (Reported by @ovizii in #57.)
• Web server refuses to boot unactivated — launching the web server directly (scripts, REPL, tests) bypassed the plugin entry's activation gate and created ~/.opencode-mem0 databases unconditionally. The server entry now guards too: unconfigured launch throws and writes nothing.

Full Changelog: v2.21.0...v2.22.0