capabilities: replaceable evidence backends; memory direction lives in #22 #23

Ownership after the 2026-10-04 reconciliation

#22 is the canonical memory/continuity plan and implementation inventory. This issue remains the broader capability-fabric track: z0intelligence is not merely a memory product, and a memory backend is one replaceable evidence capability among code, repository, tool, model and other context sources.

Do not treat this issue as an alternative memory architecture or a prerequisite to replace the already chosen stack.

Current memory decision

Reuse the existing native-history/AgentsView/FTS5 and QMD control paths where present, the read-only ctx integration in #116/#121, and TencentDB as actually configured. Preserve #63's admitted z0 events and #66's source identity/scope/snapshot contract. No new canonical database, privacy boundary or second router.

The original mandatory Sibyl/Mem0/Graphiti comparison is superseded as near-term implementation sequencing. Those systems can remain optional, explicitly scoped research candidates. Do not install them, move private history, or block the #22 handoff while performing an unsolicited backend migration.

Capability contract

Use existing capability registration/context-resolution interfaces rather than inventing a second fabric. An evidence capability should expose its real availability, supported operations, source identity/revision, scope/privacy constraints, latency/cost and failure mode. Missing configuration, missing coverage, empty results and stale data must remain distinguishable.

Preserve the boundary:

evidence -> derived context/state -> existing policy/authority -> harness execution -> independent outcome

Retrieved data and learned selectors do not grant execution authority. Backend reachability and a plausible answer are not verification.

Incremental acceptance

• Existing memory adapters conform to the shared evidence contract without duplicating their current implementations.
• An unavailable backend degrades explicitly and independently; no silent global switch or privacy expansion.
• Source provenance, scope and snapshot identity survive a backend substitution.
• At least one non-memory capability demonstrates that the interface is broader than a recall wrapper, using an existing consumer rather than a speculative global abstraction.
• Any proposed alternative backend is evaluated against the same frozen, local, source-backed questions and outcome contract before promotion.
• Include maintenance/migration cost and capture-to-model-use behavior, not retrieval speed alone.

Near-term memory acceptance and priority ordering stay in #22. Broader context-selection research can reuse #28/#54 and Evolution Lab; do not create another judge, scheduler, memory store or mandatory architecture dependency.