ampack: signed, declaratively traversable memory packs for AI agents, with trial-gated activation

Agent hosts can import memory produced elsewhere, but a verified publisher signature binds listed content to a trusted signing key; it does not establish whether that content will help the receiving host. I present ampack, a memory-pack format and admission procedure combining content verification, bounded retrieval and receiver-controlled activation. A canonical file list gives each pack a SHA-256 identity across supported representations. Signatures are checked under the receiver's trust policy. Declared relations specify graph traversal; the host enforces budgets and constructs reason paths. Under an enforced trial-gating policy, a pack contributes only with an accepted, content-bound admission decision. The receiver can evaluate it on its own history or recorded questions, check declared targets and sign the decision. Three separately implemented readers agreed with the expected verdicts on 513 fresh conformance cases with available inputs in a 5 November check; all share one author. Historical code-map trials without hindsight kept the pack inactive on four public repositories and my repository. Two synthetic knowledge packs passed observed hit and abstention thresholds under meaning-based search with a judge, finding a correct note on 32 of 40 answerable questions each. Reranked ordinary memories matched or exceeded those counts. These observations support packaging and admission enforcement in the recorded settings, not a general retrieval advantage. Golden-arithmetic checks pass, while statistical validation remains limited to specified finite regimes; enforced retrieval-configuration binding and broader validation remain limitations. Code and supporting evidence are planned for release at https://github.com/joslat/agent-memory-dotnet after publication of the preprint.