Shared AI Memory

Once several sessions share text memory, the work is data consistency: whether the index grows too large, whether concurrent writes overwrite each other, whether links break and whether old conclusions still hold.

Keep the always-loaded index to short pointers into topic files and read the details on demand. Before writing, re-read the latest version and locate an entry by file path or a stable identifier, never by a line number read earlier. Deleting an index entry does not delete its topic file, and a file that still exists may no longer be found, so check both the pointer and the target when tidying, and make sure the reader understands the stored format.

Old memories are leads, not facts: model versions, deployment state and operating policy go stale fastest, so verify them against current code, rules or live data before use, and confirm any checker fires on a known-bad sample before trusting its green result. This is maintenance advice, not an automated audit service BluffKing runs; the old memory-audit and automatic sweep workflows are retired.