Memory-CL.

Built a deterministic AI memory engine that turns a codebase into a queryable knowledge surface and exposes it to agents, developers, and operators through five surfaces (HTTP, MCP, a Python SDK, the `memcl` CLI, and a Next.js transparency UI) all correlated by a shared `X-Request-ID`.

Hybrid retrieval blends graph (Neo4j), vector (Qdrant), and metadata (Postgres) channels via a fixed-weight ranking formula; every ranked entry surfaces its score breakdown and full pipeline trace.

Determinism is a hard invariant: snapshot plus replay produces byte-identical outputs across runs, and a hash-chained governance ledger makes the audit trail tamper-evident. Production hardening includes a multi-stage Dockerfile (non-root uid 1000, tini PID 1, lockfile-strict installs), strict env validation that rejects sentinel passwords and unsafe defaults at boot, and discrete safe-mode states (`read_only` / `mcp_disabled` / `retrieval_only`) for graceful degradation.