bitter log similar --text "same failure, better evidence"

The evidence ledger for agent runs.

BitterLog gives operators one append-only evidence and memory layer for agent-built software. Preserve raw traces, search failures, and explain what actually happened without stitching together a second analytics stack.

Every run keeps stable digests, timestamps, receipts, and annotations so your team and your buyer can inspect the same evidence. During beta, onboarding and support route through a human BitterDesk thread.

See the evidence flow Request beta access

API and CLI ingest
Post operating events, upload raw traces, or bring the receipts your agent stack already emits.
Receipts stay intact
Prompts, stdout, stderr, manifests, and annotations keep stable digests and timestamps you can inspect later.
Buyer-safe beta
No ad pixels, no resale, and BitterDesk onboarding from first sample through first recurring run.

evidence ledger

14:03:12events ingestedPOST /api/v1/events batch=12

14:03:15receipt storedagent_runs / builder / 5468

14:03:22trace fetchedbucket object 71 body-only

14:03:30similarity hitrepeated git status before failure

events, receipts, traces, annotations

01 / CLI

Start with the trace. Then move outward.

BitterLog is optimized for the way agent systems actually behave: lots of raw logs, repeated research, and constant need to pivot from a summary back to the original evidence.

Ingest raw evidence

bitter log put --bucket agent_runs --run-id 5468 --agent builder tmp/agent-runs/builder_5468.jsonl

Preserve the original trace first. Parse later.

Inspect and pivot

bitter log list --bucket agent_runs --run-id 5468
bitter log get --bucket agent_runs --id 71 --body-only

Move from run identity to the exact bytes that explain what happened.

Research and explain

bitter log patterns
bitter log query --pattern recent_run_pressure
bitter log similar --text "repeated git status before failure"

Use canned patterns, read-only SQL, and fuzzy similarity over projected evidence.

02 / Identity

Logging built for agents, not generic app noise.

Raw agent logs

Session JSONL, stdout, stderr, prompt snapshots, rendered output, and receipts stay intact with stable digests and timestamps.

Search and analysis

List objects, inspect schema, run research patterns, issue read-only SQL, and explain how the analytical layer is answering.

Learning surface

Attach annotations, compare similar traces, and turn recurring agent behavior into evidence you can act on.

03 / Trust

From run receipt to reviewable proof.

BitterLog keeps custody, identity, and review separate from downstream analysis. Raw evidence remains the durable ledger while query and similarity paths explain what happened without rewriting the source record.

No custody handoff

Raw traces stay in BitterLog with stable digests and timestamps. Research surfaces work from refs and projections, not a second source of truth.

Reviewable receipts

Each run can connect request context, stdout, stderr, annotations, and deploy receipts so operators and buyers audit the same path.

Read-only query path

ClickHouse-backed search, SQL, and similarity explain patterns over projected evidence without mutating the append-only ledger.

04 / Research

Search, SQL, and similarity in one surface.

Use BitterLog for discovery, direct querying, and fuzzy retrieval over projected trace lines and annotations. ClickHouse sits behind the service for structured analysis, but the agent interface stays simple.

starter moves

bitter log patterns
bitter log tables
bitter log schema --table agent_run_summary
bitter log query --sql "SELECT * FROM agent_run_summary LIMIT 5"
bitter log explain --sql "SELECT count() FROM semantic_chunks_current"
bitter log similar --text "retried the same failing tool call"

bitter log tables

bucketagent_runs

identityrun_id, agent, origin, digest, observed_at

analyticsClickHouse on tap for read-only query paths

learningannotations plus qualitative similarity over traces

05 / Access

Bring one run, receipt, or event stream.

Tell us which traces, receipts, or operating events you want BitterLog to preserve, search, and explain. We will review the first sample with you in BitterDesk before you wire a recurring feed.