mirror of
https://github.com/supermemoryai/claude-supermemory.git
synced 2026-09-14 14:58:03 +08:00
Dhravya Shah
ed39afa764
The recall hook can only inject what it finds. On a miss it returned silently, and silence reads to the model as "there is nothing to search" — so it never reached for the tool, and history living in another container stayed invisible for the whole session. The guidance that would have fixed this already existed, but only inside formatRecall(), which runs when recall already found something. It was delivered in the case where it was least needed and withheld in the case where it mattered most. Emit a discovery block on the first genuinely-empty recall of a session: name the tool, point at ToolSearch to resolve it (the tool ships under three names depending on install shape, and deferred loading means the model holds a bare name with no schema), interpolate the repo containerTag, and note that read-only calls are auto-approved. Gated to fire once per session, and only when the container is truly empty rather than all-repeats — misses are the common case in an empty project, so repeating it would tax every prompt and invite searches on turns where memory is irrelevant. The marker is a file in the session dir next to recalled.json; writeState would have meant widening the EVENT_NAMES allowlist and touching the statusline renderer. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011qEtmi9chBJMgirVCTpSLu