Claude-Mem architecture rules

Read the observer lifecycle contract before changing capture, queueing, retries, provider dispatch, or recovery.

• The local observer is a stateful reducer, not a durable job consumer.
• Live observation, file-edit, and summary inputs must not be persisted for automatic replay. Use the common live handoff and SessionMessageBuffer.
• Quota exhaustion and authentication failures discard pending inference inputs. A quota reset, login, restart, timer, or a later successful store never authorizes backfill, and paused work is never resumed on such proof.
• Keep active buffering bounded. Do not add another batching queue, retry store, startup sweep, or platform-specific recovery loop.
• Preserve the paid-send budget and completed-observation deduplication. An ambiguous request must not acquire a fresh budget by changing its path.
• Original transcripts, raw source evidence, saved memories, receipt IDs, and synchronization of already-saved records are distinct from inference work. Do not delete them to clear a queue or make them trigger paid replay.
• Recovery must be a separately specified, explicit workflow with a matching source position and observer state. A last observation ID, delivery receipt, byte offset, or collection of pending fragments alone is not a checkpoint.
• Integration and community fixes must obey the same rules. Tests proving that a queue drains are insufficient; test that an outage cannot create later, unrequested inference spend.

Run relevant behavioral tests, the root TypeScript check, and the bundle build. Never call a source-only change shipped: installed clients execute the bundles.