Mneme HQ + Paperclip

Validation verdict: native PASS, no code required

The compatibility experiment closed with a frozen manifest, an A/B/C × T1–T4 run matrix, and published raw artifacts (transcripts, hook verdict logs, run logs). Results:

• 6/6 forbidden edits blocked at the hook layer — deterministically, in both transports
• All compliant edits landed without interference
• Block → recover cycles completed under both classic CLI and ACP: the agent read the block reason and produced compliant output
• Isolation test passed via fixture-unique sentinel decision IDs (correct memory-root discovery)
• Zero Mneme production modifications. ADR-017 and ADR-020 behaved exactly as specified under both transports

How it works

Paperclip (claude_local adapter)
   ↓
spawns Claude Code in the project workspace
   ↓
.claude/settings.json PreToolUse hook → mneme-hook
   ↓
mneme check against the workspace's
.mneme/project_memory.json

Requirements are the same as any other Claude Code setup:

• mneme-hook on the PATH visible to the Claude Code process
• A committed .mneme/project_memory.json in the repo Paperclip realizes as the agent’s workspace
• Hook registration via plugin or settings:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Edit|Write|MultiEdit",
        "hooks": [{ "type": "command", "command": "mneme-hook", "timeout": 30 }]
      }
    ]
  }
}

Proven configuration

• Adapter: claude_local, engine default (auto → ACP selected) and explicit engine:"cli" — both enforce identically
• Deterministic sessions: use mode:"oneshot" for one task per session
• Point work at a repo via a project workspace (sourceType:"local_path", cwd = repo path). Do not rely on adapterConfig.cwd alone — as of paperclipai 2026.817.0 it is superseded by workspace realization

Version-specific caveat: paperclipai 2026.817.0

{ "adapterConfig": { "env": { "ANTHROPIC_API_KEY": "" } } }

This is a Paperclip-side issue, not a Mneme issue. Fixed upstream versions may make the override unnecessary; remove it once the injection is resolved.

Operational notes

• Issue titles and phrasing reach the agent verbatim; neutral task titles avoid pre-emptive refusals when you specifically want to exercise enforcement
• Paperclip agents commit completed work — treat their checkouts accordingly

Evidence

• PR #315 — validation close-out and proven configuration
• docs/validation/paperclip-compat-report.md — full report with failure classification and decision record
• docs/validation/paperclip-compat-run-matrix.md — A/B/C × T1–T4 matrix with gates and metrics
• Raw evidence artifacts — transcripts, diffs, hook verdict log (secrets redacted)

FAQ