Agents that learn
from every run.

Captures resolved agent work — from any team, any agent — and surfaces it on the next similar run. Five failure modes it closes: repeat reasoning, forgotten file meaning, doom-loops, redundant tool calls, and context-window thrashing on long sessions.

Those are general LLM memory layers — flat retrieval over chat history. Tracebase is built for tool-driven coding agents: a pattern DB tuned for code-shape retrieval, doom-loop detection, tool-call dedup, and context folding. Vector recall is one of the five arms, not the whole thing.

No. Self-hosted by default — the pattern DB is a project-local SQLite file. The hosted dashboard is optional and read-only over the same local event log. MIT license, same binary on-prem and in CI.

One command. `npx tracebase-ai init` auto-detects the active agent (Claude Code, Cursor, Codex), writes the adapter, and registers the MCP entry. Verify with `npx tracebase-ai doctor`. Under a minute on a clean install.

The agent decides. Tracebase suggests — never overrides. Patterns that stop earning their keep get demoted automatically: the lifecycle repair loop watches outcome attribution and downranks reasoning that disproves itself.

No. Retrieval is a local SQLite read — fast enough that it isn't the bottleneck. The slower path is the agent's own work, which Tracebase shortens by skipping re-derivation.

Every retrieval, injection, agent_used, and outcome event is logged locally. `npx tracebase-ai impact` shows the 30-day funnel; the dashboard reads the same events. Injected vs used vs resolved, fully transparent — no magic.

Free while we're in beta. Self-hosted is MIT-licensed and available today; paid tiers (Hobby, Startup, Enterprise) are still in draft and not on checkout yet. Talk to us for early access.