Total-Recall — Full-System Integration Audit
Professional audit · 2026-07-24 · cost of integration to the full system vs the laboratory alpha. Every figure labeled MEASURED (verified on disk) / STRUCTURAL (code-inspection) / SPECULATIVE (stretch).
3-agent parallel auditground-truth, no fabricationgrounded-vs-speculative labeled
0Executive summary — the headline
⚠️ CORRECTED against the LIVE tree (`.deploy-main`, not the working checkout). The deployed recall is further along than a first read of this repo shows: rankRecall already runs ARK union_conf (coverage-level SELECT-not-FUSE) + L2 Grassmann + PPR-field re-rank LIVE over a 65 MB captured store (plist flags ON). So the ranking stack is largely built AND substantially live. But three gates remain to finishing the CORE system, and one is more honest than flattering: (1) the composition frontier is unbuilt — the endpoint-base producer (partial recall from Marvel/Gemma) + the content-regime SELECT router (net-new); (2) retrieval is un-instrumented (real counts/latency not measured); (3) CAPTURE-LIVE ≠ RECALL-PROVEN — the "2,390 captures" is a capture count, the graph is empty/sparse (0 edges), and the canonical ledger marks current real-corpus runs "INCONCLUSIVE-STORE-FILL" (not yet dense enough to validate recall). The genuine #1 build is the content-regime SELECT router (NET-NEW, not on main — only a research harness) + closing the credit→learned-router loop, both gated on real-data validation. Honest surprises: "Hermes memory" is a health-check STUB with no store (L-tier build); "capture is live" is deploy-tree-scoped. Operational rule: build off origin/main, NOT this RL branch (it lacks the L1–L6 plane stack). Effort: QUICK wins ~1 week · content-regime router + first conclusive validation ~weeks–months · full integration months.
1What's readied to USE the memory system (coverage)
| Subsystem | Memory status | Evidence |
|---|
| L0 · Memory (content store) | WIRED-LIVE | total-recall store, 65 MB, live provider (.deploy-main); capture live (Stop-hook→drainer→store.put) |
| Addressing · Signatures (snowflake-merkle spine) | BUILT, unwired | snowflake-spine.ts (composeCrossLayerRoot/verify); the live store addresses by sha256 + FTS, NOT the spine |
| Thinking · Ranking (coverage SELECT-not-FUSE) | WIRED-LIVE | rankRecall = ARK union_conf coverage-level SELECT, live in prod |
| Intuition (L2 PPR field) | LIVE but degenerate | PPR-field re-rank + L2 Grassmann run live — but the graph is empty (0 edges), so there's nothing to rank over yet |
| Perception (L3 / L4) | BUILT, flag-OFF | perception.ts + FORGE_PLANE_L3/L4 — integrated in rankRecall, gated on real-data validation |
| Executive · Credit (L6) | BUILT, no consumer | credit.ts + recall-router-policy.ts — loop open at the last hop (router doesn't read the policy) + starved (no real outcomes) |
| Composition (endpoint base ∪ our layers) | PARTIAL | local-plane compose is LIVE; the endpoint-base (fetchBase) producer is ABSENT, and the content-regime SELECT router is NOT BUILT (net-new, W5, research harness only) |
| Ingest / capture (the fill loop) | WIRED-LIVE | the ONE external touchpoint = api-bridge (recall context + capture-back) + the Stop-hook drainer |
Readiness across the CORE brain stack (the scope = the total-recall memory/brain we're finishing, NOT other Forge products): layers are ~90% built and substantially wired-live — L0 memory + coverage-SELECT ranking + L2/PPR run in prod over a 65 MB live-captured store — but not yet validated: the co-access graph is degenerate (0 edges) so intuition/PPR has nothing to rank yet, recall is "INCONCLUSIVE-STORE-FILL", and the composition's two frontier pieces (the endpoint-base producer + the content-regime SELECT router) are unbuilt. The endpoint (Marvel/Gemma 34.70.203.14) is IN scope — it's the partial-recall base + a core model; capturing ALL active sessions + all models (the amnesia fix) is the fill loop. Scope = core total-recall system only. Making other Forge services (Hermes/RMT/cost-tracker/…) consume it is a separate "later / data-point" question, excluded.
2Timing + count data — retrievals vs non-exposed dips
Honest headline: retrieval is un-instrumented — "how many retrievals, how fast, vs how many bypasses" is not measured today. A quoted runtime ratio would be invented.
What's REAL (persisted): the ONLY per-retrieval telemetry is a
default-OFF 201-row co-access spool (2 opportunistic sessions). Live store =
2,713 records, but
0 vectorized · 0 graph edges · 0 provenance → the relational/PPR layer is empty; recall is FTS/anchor-only. Admission-ledger = 3,042 rows, all admits
terminal=quarantined (not organic store — flagged for review).
What's NOT counted: the router's
store.find() has zero timing/logging/counter; recall latency is written to
stderr and discarded;
events.jsonl (80k lines) has
zero memory-retrieval events; dispatch has no duration field. Direct CXDB/git/fs dips are also uncounted.
Ad-hoc timing that IS real (this session): live store find() ~1.4–28ms; synthetic sweep point-lookup
sub-ms & flat to 1M, broad scan O(N)→~1.3s@1M; concurrency knee K≈8; event-loop HOL 85µs→65ms (fixed).
MEASURED (audit-timing, read-only).
One-line fixes (close the whole gap): wrap store.find() → append {q,limit,hits,us,ts} to a recall-log (always-on count+latency); appendFileSync in the CXDB reader + git-exec helper → a dips.jsonl; add durationMs to the dispatch.completed emit. All are single lines against existing seams.
3Cost of integration — laboratory alpha → full system
Where alpha is (grounded, disk-verified): library built ~90% · wired-live ~30% · real-data-validated ~15–20%. Unit cost of wiring a subsystem = ~40 LOC (the api-bridge template: loopback fetch to :3200, no auth). Two ops footguns (grounded §18): the live router serves from .deploy-main which LAGS origin/main — setting a flag on the stale deploy is a NO-OP (must ff + pnpm install + restart first); and the Modulum Bearer token must be rotated (was cleartext in git).
| Tier | Work | Effort / risk | Unlocks |
|---|
| QUICK | Fix amnesia + make recall testable: 3 one-line instrumentation hooks (count+latency) · widen capture to ALL active sessions + all models (every model's turns → recall corpus, not just Claude) · build the co-access graph (edges) so intuition/PPR stops being degenerate | ~1 week / Low | real telemetry + a non-degenerate store the live rankers can actually recall over |
| MEDIUM | Close partial → total: BUILD the fetchBase producer (fetch partial recall from the Marvel/Gemma endpoint — interface present, impl ABSENT) + BUILD the content-regime SELECT router (net-new, W5) + call it from router.search + close the credit→learned-router loop (write the missing consumer). THEN the long pole: first conclusive real-data recall validation (NC-clean, not "INCONCLUSIVE"). Rotate the endpoint token; ff .deploy-main. | ~weeks / Med (research risk: does endpoint ∪ our-layers beat our-layers-alone on real data?) | partial (endpoint) ∪ our layers = TOTAL recall, self-improving via credit |
| LARGE | Finish the deep stack: perception L3/L4 validated + enabled · snowflake-merkle spine as live addressing · advanced planes (e8 / vsa) NC-clean · edge-deployable · scale the reconstruction/crown claims to made-real | Months / High (research risk dominates — several planes are mechanism-proven-only / synthetic-scale-only §17) | the full verifiable cognitive stack — the moat |
4Transformation potential + stretch goals
The finished core system (the goal): when it captures ALL active sessions + all models and composes the Marvel/Gemma endpoint's partial recall (~60–80%) with our WICK layers (SELECT-not-FUSE), the recall-amnesia class disappears — every session and every model inherits the collective memory instead of re-deriving, the credit loop tunes recall from real outcomes, and the endpoint's partial recall becomes verifiable TOTAL recall. That is the substrate we're finishing: turn any forgetful model's partial recall into total, verifiable recall. Expanding OTHER Forge products to consume it is a separate "later / data-point" question, not part of finishing the core.
Grounded next step (today): the QUICK tier is the highest-leverage and lowest-risk — it's mostly wiring already-written code + 3 instrumentation lines + a token rotation. That alone moves the system from "captures sessions" to "actually composes total recall AND measures itself." This audit is a living document; the composition-wire (FORGE_RECALL_COMPOSITION_WIRE_v1) is already at SEED.