What is agent product memory?

Coding agents in 2026 have a strange blind spot. They can read your entire codebase in seconds, but they have no idea why any of it exists. Ask an agent to “improve the audit log” and it will produce something plausible — pagination, maybe export, maybe a filter bar — with no way to know that three enterprise customers asked for exactly one thing (filter by actor), that your April roadmap explicitly deferred it, and that the deal blocked on it renews next month.

That gap has a name now: agent product memory — the customer-reality half of your agent's context, sitting next to the code-reality half it already has.

What goes into product memory?

Four kinds of raw signal, ingested continuously and kept queryable:

• Conversations — Slack threads, support tickets, community posts.
• Calls — sales and research call transcripts (.vtt straight from your recorder).
• Notes — the founder note typed at 2am, the hallway remark that changed scope.
• Decisions — what you already chose, so new work can be checked against it.

The memory layer embeds all of it, deduplicates it, clusters it into themes, and — critically — keeps a pointer from every synthesized claim back to the raw signal that produced it.

Why do coding agents need it?

Because the failure mode of a capable agent isn't bad code — it's correct code for the wrong spec. In our own testing, an ungrounded agent asked to add audit-log filtering produced a generic filter toolbar: five filter types, none of which matched what the paying customer described, and it silently reversed a roadmap decision no one told it about. The grounded run produced a plan with one filter (by actor), a fence forbidding the other four, and a gate that stopped work until a human approved overriding the April decision.

How does it work in practice?

$ scriptonia add slack.txt call.vtt          # feed signal (free)
$ scriptonia plan "filter audit log by actor"  # → PLAN.md, every claim cited
$ claude "execute PLAN.md"                     # agent ships the PR

The output is a plain markdown file. No proprietary format, no lock-in — a PLAN.md any agent can execute and any human can review in 60 seconds.

The test for real product memory

Three properties separate product memory from a notes folder or a RAG pipeline:

• Sourced — every claim carries the id of the ticket/call/thread it came from.
• Contradiction-checked — new work is gated against prior decisions, and the gate is ⚠ UNRESOLVED until a human clears it.
• Executable — the output is a plan an agent runs, not a report a human re-types.

If any of the three is missing, the agent is still guessing — just with more paragraphs of context to guess from.