Capture continuously
A supervised, event-driven stream follows complete local transcript appends. Durable byte cursors resume without rescanning settled history.
Local-first agent memory
Every agent conversation.
One queryable lake.
Transcript Lake turns conversations from Claude Code, Codex, OMP, Droid, Kimi, and Tama hooks into one privacy-masked local archive—while they happen.
ClaudeCodexOMPDroidKimiTama
MASKLAKENDJSON
SQL SELECT evidence
06runtime inputs
01canonical schema
0cloud credentials
∞local memory
Why Transcript Lake
Your team already solved it.
Now the answer is findable.
Mask before storage
Credential-shaped and high-entropy text is replaced before an event reaches the canonical partition or any downstream projection.
Query one contract
Provider-neutral events make sessions, tools, tokens, hook decisions, labels, and cross-runtime signals available through DuckDB.
Keep authority local
Vendor stores remain read-only. NDJSON stays authoritative. Parquet and Oko projections are explicit, rebuildable derived state.
A narrow trust boundary
Raw on one side.
Masked on the other.
Transcript Lake gives every downstream consumer the same normalized evidence without making each one handle raw vendor files, secrets, or incompatible schemas.
Read the masking guarantees$ transcript-lake stream stream start roots=6 adapter runtime=claude complete_lines=7 mask assignments=2 tokens=1 entropy=1 commit files=1 failures=0 ms=50 cursor durable offset=14592
Documentation
From first stream
to recovery.
Explore 42 complete guides, references, concepts, and executable examples covering the full development-channel contract.
Browse all documentation