SurfaceWeb · Observability ScopeRead-only by default RetentionLogs export as JSONL

See every packet your agents pulled.

Console is the window into the memory layer. Every query that crosses L1–L5, every packet your agents consumed, every refactor that changed a claim — logged, inspectable, and reconcilable against your bill. The protocol promises auditability; this is where you exercise it.

The query stream

Every query, and the layer that resolved it.

The stream is the front page of Console. Each row is one query through the retrieval stack — which layer answered, how long it took, how many tokens the packet cost, and the exact model version that produced it. Most queries never reach L5; the stream makes that visible at a glance.

console.memrails.dev — query stream live · 1,000,000 / mo
Stream
Inspector
Refactors
Audit
Evals
TimeQueryLayerLatencyTokensConf.
14:02:11 how do we handle auth token refresh? L1 4ms
14:02:09 kv-cache eviction policy L2 9ms
14:01:58 which embedding model is canonical now? L3 240ms
14:01:52 contradictions on retry-backoff strategy L4 38ms
14:01:47 summarize what we know about rate limits L5 392ms 503 0.91
14:01:39 find the on-call escalation path L1 3ms
Resolved below L5
82%

Most queries never need the model. Cheap filters answer first.

Median packet
487tok

What an L5 synthesis actually costs your agent's context.

p95 latency
410ms

Across the whole stack, end to end, including compression.

Packet inspector

Open any packet. Trace every claim.

Click a row and Console reconstructs the packet end to end: the evidence bundle that went in, the synthesis that came out, and a provenance map showing exactly which claims shaped which spans of the answer. Nothing is a black box.

Input — Evidence bundlefrom L4
{
  "query": "deployment risks for...",
  "candidates": [
    { "claim_id": "clm_8f3a", "confidence": 0.91 },
    { "claim_id": "clm_2a1c", "confidence": 0.74 },
    // 7 candidates survived L4
  ],
  "intent": "summarize",
  "max_tokens": 600
}
Output — Compressed packetcompress-v1.4.2
{
  "packet": "...",
  "tokens": 487,
  "confidence": 0.89,
  "contradictions_surfaced": 1,
  "input_hash": "sha256:9c1f…",
  "output_hash": "sha256:4be0…"
}
Provenance — which claim shaped which span
clm_8f3a
0.42
clm_2a1c
0.28
clm_5d77
0.19

Every span of the synthesis links back to a source claim with a weight. If an agent acted on a packet, you can see precisely what evidence drove it — and challenge it.

Refactor activity

Watch the memory improve itself.

Memory maintenance is software maintenance. Console shows the refactor feed — each diff that changed a claim, the evidence behind it, the validators who reviewed it, and the stake and payout that settled the market.

refactor/kv-cache · diff
+ stake 12 · validator ✓
- confidence: 0.72
+ confidence: 0.86
- evidence: source-a
+ evidence: source-a, source-b
  contradictions: []
+ refactor_notes: |
+   Added second independent source.
+   Promoted confidence after review.
review · 2 validators · 1 dissent resolved accepted
Trigger
Action
Status
New trusted evidence
Upgrade claim, link source
Sources disagree
Add contradiction, demote
Topic too large
Split file, preserve links
review
Stale claim
Downgrade confidence
Unsourced diff
Reject, slash stake
Audit & compliance

Reconcile the invoice line by line.

Every compression is logged with an input hash, an output hash, and the model version that produced it. Finance can match our metered total against their own telemetry; security can replay any decision an agent made. Export the whole ledger as JSONL whenever you want.

Immutable log

Hashed in, hashed out.

Each packet carries the SHA-256 of its input bundle and its output. Tampering is detectable; replays are exact.

input_hash · output_hash · model_version
Billing reconciliation

Our number, your number.

Console totals packets and tokens the same way the invoice does. The calculator on Pricing and the ledger here agree to the cent.

packets × orchestration + tokens × rate
Export & exit

Nothing locked inside.

Usage and settlement logs export as JSONL in one operation. Memory is Markdown and Git already; the audit trail leaves with you too.

GET /v1/audit/export → logs.jsonl
audit/2026-05 · ledger
● verified · 0 gaps
{"ts":"2026-05-22T14:02:04Z","packet":"pkt_8f3a","in":"sha256:9c1f…","out":"sha256:4be0…","ver":"compress-v1.4.2","tokens":487,"billed":0.0005}
{"ts":"2026-05-22T14:01:47Z","packet":"pkt_2a1c","in":"sha256:1d77…","out":"sha256:ae42…","ver":"compress-v1.4.2","tokens":503,"billed":0.0005}
Evals & drift

Quality is measured, not asserted.

Console surfaces the same metrics the benchmark suite tracks — density, evidence precision, contradiction survival, and drift. When a new Compress checkpoint ships each Monday, the eval scores land here first, with the canary window and rollback state in view.

M01 · Density
0.81

Tokens avoided per task

M02 · Evidence
0.94

Claims that are source-backed

M03 · Contradiction
0.98

Nuance survives compression

Packet fidelity — trailing 12 weeks
▲ +0.04
0.830.87 today
Current checkpoint
compress-v1.4.2
ShippedMon 05.18
Canarypassed · 24h
Rollback< 90s armed
Driftnone flagged
06 / See it on your data

Your memory, fully observable.

Point Console at your corpus and watch the stream fill in. Read-only by default, exportable always, and reconcilable against every line of your bill.