Memory-store unification & system-prompt / persona file management (v1.1) #53

Deferred to v1.1 (turn capture + Build Context is v1.0). Needs more discussion.

Idea

Let the user curate a list of "system files" that should be unified across the
memory surfaces, with wildcards/variables for context-dependent paths — e.g.
<project_dir>/CLAUDE.md, ~/.ragger/persona/*.md. Ragger would be the source
of truth and project/serve these into each agent's native idiom.

Why it's non-trivial (substrate mismatch)

The stores are not the same shape, so "point them all at the same file" only
half-works:

• ragger = SQLite DB (memories.db) + vector search. Accessed via MCP tools
  / HTTP. Already shared by ragger CLI, OpenClaw, Hermes, and the Claude Code
  (Code tab) MCP client — these are already unified.
• Claude Code file-memory = flat Markdown (CLAUDE.md, auto-memory
  MEMORY.md), loaded whole into context at session start. No search/embeddings.
  Only Claude Code reads Markdown memory; OpenClaw/Hermes do not.
• Claude Desktop Chat/Cowork = connectors only (model-discretion tools), no
  file to link.

You can't symlink a .md to a .db; symlinks only unify Markdown↔Markdown.

Candidate approaches (the three from discussion)

1. Script (symlinks) — only for like-for-like Markdown consolidation. Minor.
2. Instruction MD for manual linking — document the hook + symlink setup. The
   manual version of (3).
3. Magic in ragger (preferred) — ragger as source of truth, projecting into
   each agent's idiom:
   • DB/MCP agents (OpenClaw/Hermes/MCP): direct, already done.
   • Claude Code file-memory: either (a) a SessionStart hook that
     live-queries ragger and emits context (no file to sync — preferred), or
     (b) a ragger export-md that renders a read-only CLAUDE.md projection.

Caution

Pick a direction to avoid loops/duplication: if turn-capture feeds ragger
and a synced CLAUDE.md feeds back into context (then gets re-captured), the two
systems will diverge. Cleanest: ragger canonical; CLAUDE.md is a derived,
read-only projection — not a second editable source.

Relationship to system-prompt management

System prompt = the agent's concern (Claude Code: CLAUDE.md / auto-memory /
--append-system-prompt / SessionStart hook; OpenClaw: systemPrompt /
*SystemContext; Hermes: system_prompt_block()). Ragger's Build Context
output is a per-turn context payload (the context channel), distinct from the
system prompt. v1.1 would add optional ragger-managed persona/system files that
project into those system-prompt channels.

🤖 Generated with Claude Code