Memory consolidation and gated reflection

Agents accumulate signal across runs — what worked, what failed, what a retrieval turned up that was actually useful. Left alone, that accumulation is drift: an agent that behaves differently next quarter for reasons no one can trace. Consolidation is how Orkena turns that accumulation into controlled improvement instead.

What consolidation does#

Overnight (or on demand), the runtime processes the memory writes a workspace’s agents produced since the last run:

• Merges duplicates — overlapping memories collapse into one
• Surfaces contradictions — where two agents (or the same agent at different times) learned conflicting guidance
• Proposes rule refinements — a candidate change the runtime observed would have improved outcomes, based on the pattern across runs

OVERNIGHT CONSOLIDATION · workspace ws_finance-ops
────────────────────────────────────────────────────────────────
MERGED           47 duplicate memories collapsed
ARCHIVED         12 stale memories retired
CONTRADICTIONS   3 surfaced for review
                 ─ refund_bot: "refunds >$5k always require approval"
                 ─ triage_bot: "refunds >$10k always require approval"
                 ─ policy: neither is the current rule
RULES PROPOSED   2 refinements queued
                 ─ pending review (24h veto window · 2 reviewers)

Nothing proposed activates#

This is the load-bearing property. A proposed rule is inert the moment it is created:

1. It sits in a veto window (configurable; 24h is the default) visible to the assigned reviewers.
2. A human reviews the proposal with its evidence trail — the runs and contradictions that produced it — and accepts or ignores it.
3. Accepted proposals promote through the same eval-gated path as any other change — see Evaluation suites. A memory-derived rule is not a shortcut around the regression suite.
4. If nobody acts within the window, the proposal expires without effect. Silence is not consent.

The system can propose; it cannot enact. That distinction is what makes an AI system’s own learning something an examiner can reason about between examinations — the runtime does not change its own behavior unilaterally.

Every read, write and consolidation is a ledger event#

Memory access is purpose-tagged: which agent read what, under which purpose, is on the record in the hash-chained ledger — so a data-governance question (“who used this customer’s data, and for what?”) answers from evidence, not recollection. Consolidation runs themselves are ledger events too, with the merge/archive/contradiction/proposal counts recorded.

Tenancy applies to learning#

Memory is scoped per workspace. What one workspace’s agents learn does not leak into another’s — the same row-level-security boundary that isolates runs and the ledger isolates memory.

Concretely#

Three agents in a claims workspace have accumulated overlapping guidance about refund thresholds. Consolidation merges the duplicates and surfaces the contradiction — agent A learned “deny above €500,” agent B “deny above €250.” It proposes one rule. A human reviews the proposal with its evidence trail and accepts or ignores it; an accepted rule promotes through the eval gate like any other change. Nothing in the loop is invisible, and nothing in the loop is automatic.