Palinode — Claude Code Integration

Palinode gives Claude Code persistent, git-versioned memory via 13 MCP tools. Memories survive across sessions, are searchable by meaning, and consolidate weekly.

Session Workflow

Before Starting Work

Use palinode_search to find any prior context on [project/topic]

Surfaces relevant decisions, people context, and insights automatically.

During Work — Save Key Decisions

When you make an architectural or product decision:

palinode_save: decided to [X] because [Y]. trade-off was [Z].
type: Decision
entities: [project/name]

After a Long Session

Save key decisions and insights from this session to palinode

---

Core Tools Quick Reference

Tool	Call Pattern
Search	palinode_search(query="...", limit=5)
Save decision	palinode_save(content="...", type="Decision", entities=["project/name"])
Save insight	palinode_save(content="...", type="Insight")
Save person	palinode_save(content="...", type="PersonMemory", entities=["person/name"])
Status	palinode_status()
Recent changes	palinode_diff(file="projects/name.md", n_commits=3)
Register trigger	palinode_trigger(action="add", description="...", memory_file="path/to/file.md")
Consolidate	palinode_consolidate(dry_run=True)

Memory Types

• Decision — what was decided and why (rationale is critical)
• Insight — generalizable lessons (applies beyond one project)
• PersonMemory — person's role, preferences, constraints
• ProjectSnapshot — current state (milestones, blockers, next steps)
• ActionItem — follow-up with owner and deadline

When to Save vs. Not

Save:

• Architectural decisions with rationale
• Bugs that took >30min to find (so future-you recognizes them)
• Performance findings (e.g., "bge-m3 cold start is 30s, warm is 100ms")
• People context (preferences, constraints, relationships)
• Lessons that apply broadly across projects

Don't save:

• Raw code (git handles it)
• Step-by-step log of what you did (save the outcome, not the journey)
• Temporary debug notes

Setup

See references/setup.md for MCP config, SSH remote setup, LaunchAgent, and troubleshooting.

Triggers

Register once, fires automatically whenever the context matches:

palinode_trigger(
    action="add",
    description="training data curation",
    memory_file="insights/curation-over-volume.md",
    threshold=0.72
)

Good patterns:

• Project-specific risks → "when working on [project], surface [file]"
• Recurring gotchas → "when deploying models, surface memory/oom-lessons.md"
• People context → "when Alice's design decisions come up, surface people/alice.md"