Cross-Agent Memory Write Discipline (Delivery)

Cross-Agent Memory Write Discipline (Delivery)

1. 01

   Name the single write authority for each client workspace before any agent starts recording

   Pick one memory layer per account so two agents never overwrite the same project note. Bourdon federates recognition across Claude, Codex, and Cursor, but federation still needs a designated origin for each fact.

2. 02

   Define the four note types your delivery team will actually write

   Decisions, constraints, debugging discoveries, and client preferences cover most retainer work. Knownbase organizes notes by project and tag, so an untyped note becomes an unfindable one three months later.

3. 03

   Attach provenance and a date to every entry at the moment of writing

   Record which agent, which session, and which client artifact produced the note. OzBrain stores structured articles with links and freshness, which is what lets a reviewer tell a stale assumption from a current one.

4. 04

   Cap each note at the decision plus its rationale, not the full transcript

   A 40-line paste of chat history dilutes retrieval. Write the conclusion, the constraint it satisfies, and the client deliverable it affects.

5. 05

   Route entity relationships into the graph rather than into prose

   Link the client, the campaign, the platform, and the owner as connected entities. Cogni retrieves through entity-graph spreading activation, so connected facts surface on multi-hop queries that a flat vector store misses.

6. 06

   Run a weekly read-back test with a fresh agent session on one live account

   Ask the agent to restate the client's brand rules and last three architectural decisions without prompting. Gaps found in the read-back become next week's write tasks.

7. 07

   Log every correction a human makes to agent-recalled context

   Corrections are the highest-value notes in the system. A wrong recall that gets fixed once should never recur on the same account.