Your agent can't fix what it never learned.

The same institutional memory that grounds your coding agents also powers autonomous PR review. Decisions, constraints, and rationale extracted from tickets, commits, review comments and merged PRs, instead of generic LLM guesses, so implementation and merge stay aligned.

└── demo.01.a — agent context

You're a senior engineer. You ask Claude to wire entitlement checks into generative search citations. Without us, the agent confidently clones infrastructure your org already has. With QuarkMemory MCP, it builds upon systems you actually run.

claude code broker citation gating
broker-citation-gating · generative search entitlements
You
User
C
Claude
CR
Code review
17 wk
Without — greenfield rebuild estimate
3 ph
With — phased extension of existing guard
60%+
Without — eng comments on factual errors first
~80%
With — product phasing debate, not scope fights
└── demo.01.b — autonomous PR review

You're a technical manager clearing blockers for a team shipping citation gating. Generic review bots miss org constraints buried in merged PR history. QuarkMemory posts line-anchored reviews grounded in decisions, constraints, break-risk, and past human review threads — so merges ship conflict-free instead of looping on facts your org already settled.

github.com/acme-search/platform-api/pull/318/files
GitHub
acme-search/platform-api
Type / to search

feat(composer): gate broker citations at render time #318

Open Changes requested alex.chen wants to merge 3 commits into main from feat/citation-gate
Changes from all commits Showing 1 changed file with +18 −3 lines
src/composer/citation_gate.py +18 −3
4141 def render_citations(self, broker_id: str, chunks: list) -> list:
4242 if not self._guard.has_seat(broker_id):
43+ log.debug(f"broker_id={broker_id} denied seat")
44+ return [] # safe mode: strip all citations
4345 return self._masking.apply(chunks)
CR
coderabbit-aiBot · inline on L43
Consider extracting the denial branch into a helper for testability. Add unit tests covering the empty-list path.
CR
coderabbit-aiBot · inline on L44
Returning an empty list is a reasonable safe default. You might document this behavior in the module docstring.
MK
marc.kim · senior eng · 2h later
Blocking — We cannot log broker_id in composer paths (PR #2291). Empty citations break compliance dashboards; degraded mode must return masked citations, not []. Bot missed both. Third review round.
Merge blocked — 0/2 bot comments cited org memory · senior eng spent review cycle on constraints already captured in merged PRs · ~3 days to green
~3 d
Without — merge blocked on rediscovered constraints
<3 m
With — webhook to published review (p50 target)
0%
Without — generic bot comments citing org memory
≥80%
With — comments grounded in facts & snapshots

Shipping requirements just got a whole lot faster.

You're a product manager. You ask for a broker research entitlements PRD. QuarkMemory frames the draft within the context of what your org has already discussed, decided, sprinted and implemented. No PM is going to manually reconcile Slack, old PRDs, or eng threads before v1 leaves your desk. We will.

// Generic LLM draft
Broker Research Entitlements PRD UNGROUNDED · ~15 sections
FR-1 · INVENTED

New entitlement data model with license expiration tracking — org already has entitlements.

FR-4 · CONTRADICTS SCOPE

Admin interface to grant/revoke broker access, bulk import, self-service view — config is out of band.

ARCHITECTURE · GREENFIELD

Dedicated entitlements microservice between search engine and citation DB. PostgreSQL + Redis + 7-year audit store.

NFR · INVENTED METRICS

Entitlement check <50ms p95. 99.9% uptime SLA. 10,000 checks/sec. Zero licensing violations.

TIMELINE

17-week greenfield build across 5 phases — eng reaction: "We already have this. Why rebuild?"

// QuarkMemory-grounded draft
Broker Research Entitlements PRD GROUNDED · 12 focused FRs
FR-1 · ORG TRUTH

Evaluate org-wide broker entitlements before broker content in generative results/citations.

FR-2 · ORG TRUTH

Resolve seat-level entitlements before generative summarization begins.

FR-3 · ORG TRUTH

Block/mask broker material when seat lacks entitlement — not after answer is composed.

FR-4 · NEW (explicit)

Sentence-level citation gating at render time — flagged in trust summary, not presented as settled fact.

NON-GOAL · ORG TRUTH

Do not replace subscription ingestion or admin entitlement configuration workflows.

TRUST SUMMARY

Grounded vs new vs needs-eng-input separated. Validation gate blocks ungrounded precise metrics before "ready to share."

~15
Sections in generic draft incl. API appendix
12
Focused FRs in grounded draft (~2 pages)
High
Eng back-and-forth on facts vs product choices
Low
Facts aligned on first review — phasing debate only

It doesn't matter what your role or task is in your enterprise. We're here to make it better.

See us on your stack

Memory that ships with your agents.

Decision lineage for code. Org truth for requirements. One graph, every persona.