Memory is curated knowledge, not an unlimited transcript
A chat history records a conversation. Product memory preserves the pieces that should inform later work: accepted goals, evidence, constraints, decisions, and lessons. An AI summary can help prepare those records, but it does not establish that they are correct or still relevant.
In the PM Harness approach, memory is deliberately maintained. A rejected proposal should remain discoverable as a rejected proposal, not resurface later as a current requirement. “The team discussed it” and “the team committed to it” are different claims.
AI product memory is a concept discussed here, not an available persistence service from yourproduct.chat.
What a useful memory record contains
- Statement: one understandable fact, assumption, constraint, or decision.
- Type and status: for example, observed evidence, proposed assumption, or accepted decision.
- Provenance: where the statement came from and what the source actually supports.
- Scope: which users, initiative, version, or period it applies to.
- Owner: the person responsible for clarifying or reviewing it.
- Validity: last review, expiry where appropriate, and events that trigger reconsideration.
- Relationships: the decision or specification that depends on it.
A short record with a clear source is often more useful than a large context dump. Do not store private material merely because it might be useful someday.
Example: preserve a constraint without freezing it forever
Illustrative situation: a team is evaluating external sharing of reports. An owner confirms that the first experiment is limited to invited workspace members. A useful record might read:
ID: M-12
Type: Accepted constraint
Statement: The report-sharing experiment is limited to invited members.
Scope: Initial report-sharing experiment, version 1
Source: Decision D-7, reviewed by the product owner
Owner: Report-sharing lead
Status: Current
Review trigger: A proposal to introduce public or anonymous sharing
Dependent artifact: Report-sharing PRD, access requirementsIn a later conversation, an assistant proposing public links should retrieve this constraint and surface the conflict. It should not treat an older brainstorming note about public sharing as approval.
If the team later accepts public sharing, retain the old record with a superseded status and link the new decision. Erasing it would make earlier specifications harder to understand.
Retrieve context for a question, not for its own sake
Start with the question being answered: “Can a report be viewed without joining the workspace?” Retrieve the current access constraint, the decision behind it, and any unresolved proposal to change it. Bringing in every historic sharing discussion can reintroduce ambiguity.
Ask the reviewer to confirm whether the retrieved material is relevant, current, authorized for use, and sufficiently complete. If two records conflict, show the conflict and request resolution rather than selecting the more recent-looking paragraph automatically.
Structured product state provides a current accepted view. A decision log explains how that view came about. Memory includes both, along with the evidence and lessons that remain useful.
Maintain memory and measure whether it helps
- At the end of a decision, save only the durable reasoning and the evidence needed to inspect it.
- Before a new initiative, review the records likely to constrain it.
- When a rule changes, identify dependent specifications and notify their owners.
- Retire stale records with a clear status; remove material that should no longer be retained.
Test usefulness with concrete retrieval tasks: can a new reviewer find why a feature was deferred, identify the current rule, and distinguish an assumption from a fact? Count unresolved conflicts and stale records as maintenance signals, not just the number of saved notes.
Use an organized product workspace to make records findable, and PM Harness to connect them to decisions and outcomes.