Ecosystem review / Memory / Context / Prompt
Memory / Context / Prompt: review and next steps
Memory should help an agent retain useful information without changing the authority of instructions or crossing tenant and task boundaries. Retrieval and compaction also need measurable quality and token budgets.
What the evidence supports
DzupAgent main observed during this review includes cross-layer memory security tests, opt-in PostgreSQL record expiry, fixed-transcript compression measurement and retrieval-economy evaluation tests. These are newer than the supplied summary; their presence is verified source evidence, not fresh consumer or live-database acceptance.
Reevaluation: Recast retention and evaluated retrieval work as reuse and consumer qualification. Do not rebuild the just-landed expiry or evaluation harnesses, and do not infer production behavior from their source presence.
The next useful outcome
A real consumer preserves trusted constraints, rejects instruction impersonation in retrieved data, enforces scope and retention, and stays within a measured retrieval/context budget.
Separate workstream plan
MC-01 · Proposed / reuse existing work
Qualify isolation on a consumer
Reuse the landed cross-layer tests and exercise the actual app boundary.
Acceptance: Cross-tenant, project and task leakage is rejected through the consumer route.
MC-02 · Proposed / reuse existing work
Verify compaction authority
Exercise older trusted constraints alongside newer untrusted content and compression.
Acceptance: Compaction preserves instruction authority and necessary constraints under bounded context.
MC-03 · Proposed / reuse existing work
Qualify landed expiry and recovery
Reuse opt-in PostgreSQL expiry and verify configured retention, restart and recovery on the selected backend.
Acceptance: Expired data is handled as specified and restart does not violate retention or scope.
MC-04 · Proposed / reuse existing work
Adopt retrieval evaluation
Reuse the landed retrieval and compression measurements against representative consumer data.
Acceptance: Relevance, leakage and prompt-size budgets are recorded with repeatable inputs.
MC-05 · Proposed / reuse existing work
Standardize provenance
Bind prompt versions, context snapshots, retrieved sources and cache scope to the task.
Acceptance: A result can be traced to the prompt and context actually supplied.
MC-06 · Proposed / reuse existing work
Gate durable observation by need
Add observation only for a named consumer with a bounded retention and privacy requirement.
Acceptance: The consumer benefit and scope are demonstrated before any persistent collection is enabled.
Dependencies and sequence
A named consumer, scoped storage backend, trusted-message hierarchy and repeatable evaluation fixtures.
Start with the first eligible unfinished task above, after checking its existing owner and admission. Later tasks depend on the preceding proof where applicable; this ordering does not override a current implementation plan.
Deferred work
A new memory platform, unqualified backends and durable observation without an actual consumer requirement.
Evidence and updates
This article combines the supplied 8 November ecosystem review with selected current source and retained planning records. It is an editorial snapshot, not a live service census. Detailed revision bindings and task authority remain in Workspace Docs and private execution records. Publish subsequent accepted changes as linked engineering or release notes rather than silently rewriting this baseline.
Shared roadmap · All Memory / Context / Prompt articles · Explore all areas
