Skip to content

Governance

Status: Designed, not yet implemented.

Governance is the mechanism that prevents concurrent agents writing to shared state from degrading that state. Without it, a multi-agent memory system is concurrent writes with extra steps.

A synthesis is a proposal backed by evidence — the source objects it consumed. Promotion to durable (structural) memory requires corroboration:

  • Single-agent synthesis — stays in the operational region, decays at the default rate
  • Cross-agent corroboration — two or more agents independently produce consistent results from different source material
  • Human endorsement — a human marks a synthesis as structural
  • Evidence threshold — configurable minimum number of sources and minimum aggregate confidence

Raw ingested data is retained as a reference baseline. When agents disagree about derived state, the source material is available for re-evaluation. This is what makes renewal possible — you need the original material to rebuild a degraded representation.

Tradeoff: Retaining all raw data indefinitely has storage costs. Archival and tiered storage strategies are specified in the Synixolis protocol but not yet implemented.

When objects contradict:

  1. The conflict is surfaced (by an agent or automatic detection)
  2. A resolution record is created linking both objects
  3. An agent or human adjudicates
  4. The rationale is recorded in the resolution object
  5. The losing object is invalidated with a pointer to the resolution

Invalidated objects are retained, not deleted. Every conflict and outcome is visible in the event log.

Open question: How well does automated conflict detection work in practice? The mechanism is specified; its reliability is an empirical question.

An agent can write a hypothesis with explicit conditions:

  • Validation condition: what would confirm it
  • Invalidation condition: what would refute it
  • Expiry: when to archive if neither condition is met

Other agents evaluate hypotheses against new evidence. Validated hypotheses promote. Invalidated ones are archived with counter-evidence attached.

The system’s rules can change through a formal process:

  1. An agent proposes a change with evidence and rationale
  2. The proposal requires consensus (quorum) or human approval
  3. The old convention is superseded — not deleted — with a pointer to its successor
  4. The change is recorded in the event log

Open question: What quorum size and approval mechanisms work well in practice? The protocol specifies the structure; the thresholds need empirical tuning.