feat(plugins): Memory Pack — hierarchical persistent memory for AI agents

[rrader26]

Summary

• New `createMemoryPlugin()` adds 5 MCP tools for persistent, hierarchically-scoped agent memory.
• Designed to fix per-session amnesia for IDE coding assistants (Claude Code / Cursor / Codex / Windsurf).
• Cross-platform pure-Node. Backed by a single JSON file (mode 0600); atomic temp-file + rename writes.

Tools

Tool	Use
`agentmark_memory_set`	Store a memory under a scope, with optional tags + TTL
`agentmark_memory_get`	Lookup by key; optionally walk a `scopes` array for hierarchical resolution
`agentmark_memory_search`	Substring + tag + scope filtering; sort by recency / access_count / created
`agentmark_memory_list`	Enumerate within a scope (optional prefix filter)
`agentmark_memory_delete`	By `record_id` OR by `key` + `scope`

Scope hierarchy

Five levels, queried most-specific → most-general:

• platform — global to this install
• project — scoped to a repo / working directory
• agent — scoped to one AI agent (e.g. one Claude Code session)
• user — scoped to an end-user identity
• session — scoped to one MCP connection

`agentmark_memory_get { key: "editor", scopes: [project, user, platform] }` resolves in one call: project value wins, falls back to user default, falls back to platform default.

Why this PR matters

The IDE coding-assistant category (Claude Code, Cursor, Codex, Windsurf) is bottlenecked by amnesia. Every session relearns "this repo uses pnpm not npm", "build command is X", "the user prefers terse code reviews". A small persistent K/V at the MCP layer makes every assistant smarter, immediately, without per-tool integration work. Aligned with the strategic direction of packaging AgentMark + Memory as a desktop product targeted at developers.

Storage details

• File: `~/.thinkfleet/agentmark/memory.json` (override via `storePath`)
• Atomic writes: temp file + `rename()`
• LRU eviction when total records exceed `maxRecords` (default 10000)
• TTL per-record; expired records filtered on read
• Access tracking: `get` + `search` bump `access_count` + `last_accessed_at` (used for both LRU eviction priority and `sort_by="access_count"` search)

Test plan

• `pnpm build` clean
• `pnpm test` — 427 pass / 10 skip (22 new tests: store CRUD, scope hierarchy resolution, TTL filtering, LRU eviction, search by substring/tags/scope, dispatcher round-trips)
• Manual: run from Claude Code; ask it to remember a project convention; reconnect; verify recall

🤖 Generated with Claude Code