One MCP memory file, 4 agents, and six months of bugs and stale truths

dahshan-labs · reddit · 2026-07-23

The author describes six months of running 3–4 agents on one shared MCP memory file and what broke once multiple writers started updating it.

What failed

Two writers, one file: plain overwrites silently deleted real decisions; the fix was a locked three-way merge and refusing to save if a card was lost.
Duplicates reappeared: de-dup logic kept reintroducing twins because the merge treated deleted items as if they should still exist.
Byte-identical content still conflicted: re-serializing the same card produced different bytes, which the merge misread as a conflict.
Bad eval numbers: an early 95% recall claim did not reproduce; after hardening the judge, the project saw 0/20 cold, 55% with a brief at session start, and 73% with one search round.
Stale knowledge beat corrected knowledge: similarity search often preferred a rejected proposal over the later decision that superseded it, so the system now warns when a card is stale.
Thresholds trained on fixtures failed in the wild: overlap detection that worked on short tests fell apart on the real six-month file.

What surprised the author

The memory brief stayed relatively small even with 900+ cards because it is traversed as a graph rather than replayed as a log.
Different agent sessions began coordinating through the file: lane claims, one-shot messages, and drift flags tied to git blob hashes.

Main takeaway

The hard problem in agent memory is not storage or retrieval. It is write contention and truth maintenance in a multi-writer system.

More from coding & agent