agent behavior: encode recall-first, store-last, milestone snapshots, and correction receipts

[tcconnally]

Problem

A memory server can be connected and still unused. Agents need a procedural contract for when to recall, what to persist, how to snapshot progress, how to record corrections, and what never to store.

Proposed change

Publish and implement a Perseus agent/context policy for the plugin and its integrations: recall by task nouns and project identifiers before assuming; persist durable cross-session facts, decisions, and corrections; save dated milestone snapshots before handoff; append corrections rather than rewriting history; show a compact receipt for durable changes; and never persist secrets, API keys, transient debugging noise, or repository-local facts that belong in the repo. Do not silently force writes; sensitive writes should be explicitly confirmed or governed by host policy.

Acceptance criteria

• Policy is documented for Vault MCP clients and Hermes integrations.
• Examples cover recall-first, milestone snapshot, correction, and skip-list behavior.
• Durable writes return a compact receipt with scope/key/status and no secrets.
• Tests cover secret/transient/repo-local exclusion and explicit confirmation for sensitive writes.
• Agent behavior is measured for recall/store omission, false writes, and receipt visibility.

Related existing work

• feat: emit provenance-bound context decisions with explicit abstention and stale-evidence handling #12 provenance decisions
• feat: add provider lifecycle invalidation and cache-safe forget semantics #14 lifecycle invalidation
• Vault #859
• Perseus #923 and #924

Research basis

The MemoryPlugin Field Guide documents the relevant memory admission, failure, provenance, security, retrieval, and evaluation tradeoffs:

• https://www.memoryplugin.com/wiki/rag-vs-memory.html
• https://www.memoryplugin.com/wiki/failure-modes.html
• https://www.memoryplugin.com/wiki/pipeline-anatomy.html
• https://www.memoryplugin.com/wiki/privacy-and-pitfalls.html
• https://www.memoryplugin.com/wiki/evaluating.html

This is product-research input and practitioner evidence, not an independent claim that MemoryPlugin product numbers generalize to Perseus.