A lot of teams are now giving coding agents a shared memory that follows them from one tool to another. That's the right direction.
The hard part, at least in what we measured, is not storing the memory. It's noticing when it is no longer true. In our Big Agentic Brain evals, when the only entries in memory were outdated, Sonnet saw the stale one in 88% of those runs and flagged it in only 4%. Success fell back to the no-memory baseline.
So the company brain we built treats upkeep as part of the design: a correction supersedes the old entry, and an optional valid-until date archives it the day after. Memories do not expire by default; the date is for time-bound facts.
I wrote up the broader case here: https://nevolin.be/blog/company-brain-shared-board-for-agents
The code is here: https://github.com/enhesa/big-agentic-brain
I'm involved in building BAB.
Top comments (1)
The 88%-saw / 4%-flagged split is the number that should scare people off naive memory layers. In our own setup the failure was identical: the model doesn't ignore the stale entry, it faithfully acts on it, which is worse than no memory because it's confident. Two things that helped us beyond supersession: (1) marking time-bound facts at write time instead of hoping the reader notices staleness at read time — the writer always knows more about a fact's shelf life than the reader does; and (2) a cheap recency/decay signal so a memory that hasn't been reconfirmed in N runs gets surfaced as "verify before trusting" rather than silently used. Question on your valid-until design: do you ever get bitten by a fact that's still true past its date (archived too early), and if so do you lean toward shorter dates + re-affirmation, or longer dates + explicit correction? That trade-off between false-expiry and stale-survival is the part we're still tuning.