Chat memories in B2B
Ten years in B2B product, four of them building AI. What shipping memory for a B2B agent platform taught me about trust, forgetting, and org design.
Think of a company that has no memory.
Every morning, everything starts again. The sales rep relearns which customers hate phone calls. Support rediscovers that one account always pays fifteen days late, and always pays. The same pricing debate happens every quarter, with the same conclusion, because nobody remembers reaching it. Every employee is permanently new.
No real company operates this way. An institution is, more than anything, accumulated memory: conventions, relationships, and lessons that outlive any single conversation and any single employee.
Yet this is exactly how most companies have deployed AI. Each chat begins at zero. The context you patiently explained yesterday is gone today. The correction you made last week gets unmade this week. We hired the smartest intern in the world, and we wipe its brain every night.
I have spent ten years as a B2B product manager, from 2016 to 2026. The last four were spent building AI products, and today I lead chat memory for a B2B agent platform. What follows is my field notes: two stories that changed how I build, and the principles I now refuse to ship without.
What memory actually is
Memory is what carries over between conversations.
Not the transcript. Storing every message and stuffing it back into context is not memory, any more than a security camera archive is a brain. Memory is extraction: the system reads a conversation and decides what it means. “This user wants totals as CSV.” “This buyer disputes freight charges, so check before autopay runs.” “Legal reviews any quote above fifty thousand.” The transcript is evidence. Memory is the verdict.
In B2B, that extraction applies at two levels, and the distinction matters more than any other design choice you will make.
Personal memory is applied at the user level: your preferences, your working style, your current projects. It makes the assistant feel like it knows you.
Collective memory is shared: conventions, policies, and skills, learned across all members of the company. When one person teaches the assistant how the company categorizes invoices, everyone’s assistant should know it the next morning. This is the layer that turns a chat tool into an institution. It is also the layer almost nobody builds, because it is the harder one.
A test I like: if the assistant gets smarter for one user but stays a stranger to the teammate sitting three feet away, you have built a consumer product wearing an enterprise badge. The morning after any conversation, ask what the company now knows that it did not know yesterday. If the answer is nothing, you have a chat log, not a memory.
Why Chinese companies like OpenClaw
There is a Chinese essay making the rounds on why companies there have taken so warmly to OpenClaw, the viral open source agent people run on their own machines. Beneath the memes, I think the affection has a simple root: OpenClaw treats memory as plain files. Markdown you can open, read, edit, and carry. When the agent remembers something wrong, you delete a line. When you want it to know something, you write a line.
That is not a sophisticated memory architecture. It is something better: a legible one. People trust memory they can see and touch, and are unnerved by memory that lives as an invisible blob behind an API. Hold that thought, because every expensive lesson below comes back to legibility.
A confident wrong memory is worse than forgetting
Three weeks after we shipped memory, the assistant greeted a user with a warning about a credit hold on her account. The hold had been lifted twenty days earlier. Stale state, recalled as fresh fact, delivered in the confident tone these models use for everything.
Her support ticket said it precisely: “It’s worse than forgetting. It remembers wrong and sounds sure.”
Humans forgive forgetting. We all forget, and we have social rituals for it. Nobody has a ritual for a colleague who asserts, warmly and without hesitation, something that stopped being true three weeks ago. Confident misremembering reads as lying or broken, and nothing in between. Unmanaged memory is just hallucination with a longer fuse.
The fix was not better recall. It was humility, engineered in three parts.
First, provenance. Every remembered fact now carries its source and its age. The assistant does not say “you have a credit hold.” It says the account showed a credit hold as of June 3, and anything time sensitive gets verified against the live system before it is asserted. Memory becomes a hypothesis to check, not a truth to announce.
Second, decay. Old context does not sit in the store at full strength forever. It gets hedged first (”last I knew...”), then retired. A fact about payment behavior from eight months ago should not speak with the same voice as one from last week.
Third, an edit path. When memory is wrong, the user can see why the assistant believed what it believed, and correct it in one step. The OpenClaw lesson again.
Graceful forgetting took longer to design than remembering did, and it earns more trust. If your roadmap says “add memory” on one line, split it into two: remember, and forget. Budget more time for the second.