Skip to content

Trust boundary

memd separates retrieved candidates, canonical records, and recorded verification claims. These are evidence categories inside a trusted local store. They do not authenticate the identity of an agent or reviewer.

Retrieval and canonical records

Search and digest helpers return candidates. Canonical artifacts preserve the claims, references, and provenance that a consumer can inspect. Digests are derived summaries and can become stale. artifact.find_related finds overlapping evidence; a retrieval hit does not establish a claim.

The general artifact workflow permits promotion to VerifiedRecord when a verification record supplies a different agent_id from its parent and sets supports_claim = true. Those IDs are caller supplied. This rule records a claimed independent review; it cannot establish that two distinct people or agents performed it. Consumers must assess the evidence and its origin.

Experience checks

An explicit experience check runs the supplied argv in the requesting CLI and records the exit result, timing, output hashes, and source fingerprints. A supporting receipt requires a successful exit, complete output hashes, and present, equal before/after fingerprints. This establishes only the recorded check result over the covered source.

The lower-level experience.record_check API accepts caller-reported historical receipts. Imported receipts are also reports: import does not rerun the command or authenticate its origin. Experience records always retain Canonical promotion state. Import rejects envelopes that try to attach verification or approval authority to them.

Hostnames, environment-supplied native IDs, model names, and hook fields are provenance observations. They are useful for tracing work, but are not credentials or signatures. Read the experience model for source coverage, command cleanup, and transfer limits.

Local security posture

  • The warm worker creates its Unix socket in a 0700 runtime directory and chmods the socket file to 0600 before accepting connections.
  • Embedding model and tokenizer downloads for all-MiniLM-L6-v2, BAAI/bge-small-en-v1.5, BAAI/bge-base-en-v1.5, and Qwen3-Embedding-0.6B are pinned to immutable Hugging Face commit revisions and verified against compiled-in SHA-256 digests.
  • Dense indexes record their embedding model. A different model is rejected, including one with the same vector dimension. A writer open with MEMD_BACKFILL_HNSW_ON_STARTUP=1 opts into rebuilding from segments.
  • Corrupted or tampered embedding model/tokenizer files are rejected and never loaded.
  • CLI writers share a data-directory writer lock. This coordinates processes; it does not authenticate the claims they write.
  • Explicit check commands run with the caller's local permissions. The executor is not a sandbox, and stored argv must not contain secrets.

See the task memory schema for the full canonical-artifact envelope and how trust tiers are persisted.