Skip to content

ADR-0195: One rendered recall allowance per model request

  • Status: Implementation in progress under ADR-0178
  • Date: 2026-10-12
  • Related: ADR-0183, ADR-0190, ADR-0192
  • Detailed design: docs/plan/general-memory-recall.md

Context

Automatic base and delta recall share the in-turn allowance, but each deliberate memory.search result can add another twenty items and two thousand tokens to the same model request. Comparing strategies with those different effective context caps would violate the M33 evaluation contract.

Decision

Share one twenty-item, two-thousand-token rendered allowance across retained memory.search results and automatic base/delta recall in each model request. The frozen prefix snapshot keeps its separate existing class. Corrections remain outside this allowance and must still be read when no optional recall fits.

The builtin search returns content-free counts bound to the exact rendered bytes. Only its successful result may persist that receipt through the tool pipeline. Context assembly keeps complete search result blocks in reverse conversation order, deduplicates identical blocks, and withholds whole blocks that cannot fit. At most four retained search blocks can enter dependency validation. It keeps the call/result pair, trust and canonical event content; a budget omission marker replaces only the model presentation. Older results without a trace-bound receipt are withheld in model context; pure budget inspection charges legacy blocks conservatively until validation. Changed or excerpted bytes cannot borrow an old receipt, nor can one payload borrow another payload's trace permission.

Every assembly revalidates retained results through current authoritative storage under one shared second. The trace must match its owner, session, current run, rendered digest and item count. Original records pass the same hard predicates; statements, metadata, merge membership, derived support and the actually rendered source quotes must remain unchanged. New quotes never enter this validation path. Unknown projections are withheld; the builtin memory.search produces belief packages, while automatic People context continues to be freshly assembled. Errors withhold unvalidated blocks and cancellation propagates.

Semantic-derived blocks additionally require current runtime authority and release admission. This check does not revive a query vector after checkpoint changes or reserve further search work. The search deadline limits additional work; consuming already produced evidence still checks current run/lease/transport authority and every source, even after that work allowance is spent. The trace check adds no embedding/provider call or ranking pass. Its latency belongs to assembly cost.

Automatic query preparation and context assembly derive the same remaining allowance from the checkpoint with the same current-evidence checks. Automatic retrieval uses that remainder; its existing base and delta accounting and current-source validation remain. This is deterministic context selection, not a cached answer, a source visibility grant or a global relevance rank. The evaluator must continue to distinguish per-query ranking from final visibility.

This bounds the memory evidence simultaneously shown to each reader. Repeated model input/output usage remains charged to the existing run cost/token budgets; it is not made free by a matching receipt. The four-query, two-round preparation reservation and shared deadline remain separately enforced. Complete search-work accounting, assisted planning comparisons and passing release receipts are still required.

Verification required

Regressions must cover automatic plus deliberate evidence, parallel and repeated search blocks, deferred calls, legacy results, byte/receipt mismatch, canonical event preservation, checkpoint replay, current-source/authority changes, partial source packages and corrections after allowance exhaustion. Production-shaped comparisons must use the same cap for control and candidate.