Palinode README

┌─ palinode ─┐
│ ░░░░░░░░░░ │
│ ▓▓▓▓▓▓▓▓▓▓ │
│ ██████████ │
└────────────┘

Audit-grade memory for AI agents. Every stored claim can carry an explicit epistemic status (fact / inference / open_question / unverified), typed links to the evidence that backs or contradicts it, and a verifiable quote-level citation to its source. MCP-first, so one memory works in every editor.

Agent memory is becoming a commodity. Auditable agent memory is not: Palinode is the memory layer where a remembered fact says how strongly it is believed and what evidence backs it — and where every fact is a line in a git-versioned markdown file, readable, diffable, attributable to the commit that recorded it, and revertible when it's wrong. An unmarked memory is unmarked, never silently promoted to fact. No black-box vector store you have to trust.

Your agent's memory is a folder of markdown files. Palinode indexes them with hybrid search, compacts them with an LLM, and serves them through MCP — so the same memory works in Claude Code, Cursor, Windsurf, Zed, VS Code (Continue/Cline), and any other MCP-compatible editor. Bring your own Obsidian vault, or use Palinode as one: palinode init --obsidian /path/to/vault scaffolds a full vault with graph defaults, daily-notes wiring, and an LLM-maintained wiki contract. Enterprises can govern AI memory the same way they govern code. If every service crashes, cat still works.

A palinode is a poem that retracts what was said before and says it better. That's what memory compaction does.

Built by Paul Kyle at phasespace-labs. See AUTHORS.

Supported Platforms

All platforms share the same MCP server — install once on your server, connect from any IDE. docs/HARNESSES.md is the cross-harness map: what each harness gets (native hooks vs. plugin vs. MCP), how the tiers stack, and where to start. Per-client config snippets: docs/MCP-SETUP.md and docs/MCP-INSTALL-RECIPES.md.

The Idea

Most agent memory is a black box. You can't read it, you can't diff it, you can't grep it when the vector DB is down. Palinode bets on plain files as the source of truth and builds everything else as a derived index.

Files (markdown + YAML frontmatter)
  ↓ watched
Index (SQLite-vec vectors + FTS5 keywords, single .db file)
  ↓ queried by
Interfaces (MCP server, REST API, CLI, OpenClaw plugin)
  ↓ compacted by
Consolidation (structured operations → validation and application → git commits)

That's the whole architecture. One directory of .md files, one SQLite database, one API server. No Postgres, no Redis, no cloud dependency.

One Backend, Every Interface

Palinode doesn't care how you talk to it. The full toolkit — save, search, doctor, dedup-suggest, orphan-repair, diff, blame, rollback, and more — works through every interface:

Set up once on a server. Connect from any machine, any IDE, any agent framework. The MCP server is a pure HTTP client — it holds no state, no database connection, no embedder. Point it at the API and go.

{
  "mcpServers": {
    "palinode": { "type": "http", "url": "http://your-server:6341/mcp/" }
  }
}

That's the entire client config. Works with Claude Code, Claude Desktop, Cursor, Windsurf, Zed, and VS Code (Continue/Cline). palinode-mcp-http serves streamable-HTTP at /mcp/ — use "type": "http", not "type": "sse". Always include the trailing slash in the URL. See docs/MCP-SETUP.md for editor-specific install recipes.

How It Works

Store — Typed markdown files (people, projects, decisions, insights) with YAML frontmatter. Git-versioned. Human-readable. Editable in Obsidian, VS Code, vim, or anything.

Index — A file watcher embeds with BGE-M3 and indexes with FTS5 as you save. Content-hash dedup skips re-embedding unchanged files (~90% savings). Single SQLite file, zero external services.

Search — Hybrid BM25 + vector search merged with Reciprocal Rank Fusion. The two arms have different jobs: on full-sentence questions the vector arm does nearly all the retrieval (FTS5 requires every query token to co-occur, which questions rarely satisfy), while BM25 catches the exact terms and identifiers embeddings blur. Measured together: 0.981 evidence recall@10 on LongMemEval_S (benchmarks). Optional associative entity graph and prospective triggers.

Compact — Weekly consolidation where an LLM returns structured operations and Palinode validates and applies them. Every compaction is a git commit you can review, blame, or revert.

Dream — If you've met "dreaming" as the name for this, palinode dream is an alias for palinode consolidate. Use --dry-run to inspect the proposed operations; each completed pass lands as a git commit, so a bad consolidation is a diff to review and a commit to revert.

Audit — git blame any fact. git diff any change. rollback any mistake. These aren't just git-compatible files — palinode_diff, palinode_blame, and palinode_rollback are first-class tools your agent can call.

Requirements

• Python 3.11+
• Git
• Ollama with bge-m3 (ollama pull bge-m3, ≈1.2 GB) — for semantic search. Optional: without an embedder Palinode runs in keyword-only mode (BM25/FTS5) — save, search, and audit all still work; you just don't get vector recall until you add one.

Optional extras: a chat model for weekly consolidation (any 7B+ that outputs JSON), OpenClaw for agent plugin hooks.

Install

Homebrew (macOS/Linux) — quickest path to the CLI:

brew install phasespace-labs/palinode/palinode
palinode --version

That taps and installs in one command. It puts the palinode CLI on your PATH; the service binaries (palinode-api, palinode-watcher, palinode-mcp) currently live in the tap's private prefix, so for running services use the source install below or Docker.

From source — the full-stack path. Clone, install, point at a memory directory, check it worked:

# 1. Get the code (lives separately from your memory — never the same directory)
git clone https://github.com/phasespace-labs/palinode ~/palinode-src && cd ~/palinode-src
python3 -m venv venv && source venv/bin/activate
pip install -e .

# 2. Create your memory directory (this is where your data lives — keep it private)
mkdir -p ~/.palinode && cd ~/.palinode && git init
cp ~/palinode-src/palinode.config.yaml.example palinode.config.yaml
# memory_dir stays commented out in the copied config → it inherits PALINODE_DIR below

# 3. Start the services (each in its own terminal — or as a service, next section)
PALINODE_DIR=~/.palinode palinode-api        # REST API on :6340
PALINODE_DIR=~/.palinode palinode-watcher     # auto-indexes on file save

# 4. Did it work?
palinode doctor

palinode doctor is the single "did it install correctly?" command — run it after every install, upgrade, or server move. On a fresh venv, use the venv's absolute path for the MCP command (~/palinode-src/venv/bin/palinode-mcp) so it resolves its own dependencies — see llms-install.md for the wrong-Python trap.

Your memory directory is private — it holds personal data. Never make it public; the code repo contains zero memory files. For a pre-populated demo, copy examples/sample-memory/ into ~/.palinode/.

Running as a service

Three long-running processes (API, watcher, embedder) shouldn't live in terminal tabs. Pick one:

With compose, your memory stays on the host at ~/.palinode (override with PALINODE_DATA_DIR) — the containers mount it; files remain the source of truth. Already running Ollama on the host? OLLAMA_URL=http://host.docker.internal:11434 docker compose up -d palinode-api palinode-watcher skips the bundled one. Verify any of the three the same way: palinode doctor (or curl http://127.0.0.1:6340/status).

Connect your editor

Palinode speaks MCP. Generate the correct config with the CLI instead of hand-pasting JSON (which drifts):

palinode mcp-config never writes your editor's config file — it prints a ready-to-paste block (pipe it straight to the target). Full per-harness detail: docs/MCP-INSTALL-RECIPES.md.

Daily use — drop into a project

Already installed with palinode-api running? Scaffold any project in one command:

cd your-project
palinode init

That scaffolds .claude/CLAUDE.md (memory instructions, appended if one exists), .claude/settings.json (SessionStart + SessionEnd + UserPromptSubmit hook registration), all three hook scripts, and .mcp.json (points Claude Code at the palinode MCP server). Sessions then start smart, recall as they go, and end captured: the SessionStart hook injects your core: true memories into every fresh session (startup and /clear) so standing context is there before the first prompt; the UserPromptSubmit hook recalls relevant memory before each prompt — prospective triggers plus a strict-threshold search, injected as compact snippets, silent when nothing matches; and the SessionEnd hook auto-captures on /clear, logout, and exit. Re-run with --dry-run to preview, --force to overwrite, or --no-mcp / --no-hook to scope it. See examples/hooks/ for tuning knobs.

Projects that use other harnesses get the same memory instructions automatically: when AGENTS.md (or a .agent/ directory) exists, init appends a harness-neutral memory block to AGENTS.md (read by Codex, Antigravity, and other AGENTS.md-aware agents), and when a .cursor/ directory exists it writes .cursor/rules/palinode.md for Cursor. Force or skip with --agents/--no-agents and --cursor/--no-cursor — same recall/save/session-end contract, minus the Claude-Code-only machinery (/clear, /wrap, hooks).

Usage Examples

A few common flows. Every command and option is in docs/CLI.md.

Save a decision, recall it later

# During a session — save a decision
palinode save --type Decision "Chose SQLite over Postgres for the cache layer. \
  Reason: no ops burden, single-file deployment, good enough for our scale."

# Next week — search for it
palinode search "database decision for cache"

End-of-session capture

# Agent calls at end of coding session
palinode session-end \
  --summary "Migrated auth from JWT to session tokens" \
  --decisions "Session tokens stored server-side, 24h expiry" \
  --blockers "Need to update mobile client auth flow"

Audit trail — who decided what and when

# Trace a fact back to when it was recorded
palinode blame decisions/auth-migration.md

# Compose the full provenance lineage of a fact — sources, saved/changed
# commits, supersession trail, typed links, and recall — in one view
palinode trace decisions/auth-migration.md

# See what changed across all memory in the last week
palinode diff --days 7

# Retire a memory that turned out to be wrong — it leaves recall but stays
# on disk, in git, and in the index. Never a hard delete.
palinode archive insights/stale-finding.md --reason "superseded by the re-run" \
  --superseded-by insights/corrected-finding.md

Tools

29 tools available through every interface:

Every tool is accessible as palinode_<name> via MCP, palinode <name> via CLI (hyphenated: palinode archive-expired; session_init is palinode prime; doctor_deep has no separate CLI command), or POST/GET /<name> via the REST API.

The CLI has more commands than the tool list — service control, migration, repair, and wiki-maintenance helpers. docs/CLI.md is the full command reference, one entry per command with options, defaults, and output behaviour.

Stack

Memory File Format

---
id: project-palinode
category: project
name: Palinode
core: true
status: active
entities: [person/alice]
last_updated: 2026-04-05T00:00:00Z
summary: "Persistent memory for AI agents."
canonical_question: "What is Palinode and what does it do?"
---
# Palinode

Your content here. As detailed or brief as you want.
Files marked `core: true` are always in context.
Everything else is retrieved on demand via hybrid search.
The `canonical_question` field anchors the file to the question it answers, improving search relevance.

Open in Obsidian

Palinode stores every memory as a plain markdown file — which means your memory directory is already a valid Obsidian vault. Point Obsidian at the folder and you get graph view, backlinks, and Bases on top of Palinode's hybrid search and compaction. No sync job, no plugin to install, no two-source-of-truth problem.

palinode init --obsidian ~/palinode-vault

This scaffolds the vault directory layout, an _index.md Map of Content, a _README.md orientation page, and an opinionated .obsidian/ config (graph view colour-coded by category, daily-notes wired to daily/). Then open the directory in Obsidian.

The LLM follows a wiki-maintenance contract — it keeps entities: frontmatter and [[wikilinks]] in the note body in sync so the Obsidian graph stays accurate as new memories are saved. When you save a memory with entity references, Palinode appends an idempotent ## See also block linking them as wikilinks.

Two embedding-aware tools support wiki hygiene: palinode_dedup_suggest checks whether a draft overlaps an existing file before creating a duplicate, and palinode_orphan_repair finds semantic matches for broken [[wikilinks]]. Both are callable via MCP, CLI, and REST.

See docs/OBSIDIAN.md for the comprehensive guide: quickstart, wiki contract details, migration paths, and FAQ.

Diagnose with palinode doctor

Silent misconfiguration — a db_path pointing at the wrong file, a watcher indexing a stale directory, a phantom DB file — is the most common reason Palinode doesn't behave as expected after an upgrade or server move. palinode doctor catches this entire class of bugs.

palinode doctor

The command runs 18+ checks across paths, services, config consistency, index health, and disk state, and emits a structured report with a pass/warn/fail status for each. --fix mode applies safe automated repairs (creates missing directories, appends the CLAUDE.md Palinode block) — it never moves user data; phantom DB files and DB-path mismatches print suggested mv commands but never execute them.

Run palinode doctor after every install, upgrade, or server migration. See docs/DOCTOR.md for the full check catalog and --fix reference.

Configuration

All behavior is in palinode.config.yaml:

memory_dir: "~/.palinode"
ollama_url: "http://localhost:11434"
embedding_model: "bge-m3"

search:
  hybrid_enabled: true
  hybrid_weight: 0.5         # 0.0 = vector only, 1.0 = BM25 only

consolidation:
  llm_model: "llama3.1:8b"   # any chat model that outputs JSON
  llm_url: "http://localhost:11434"
  llm_fallbacks:              # tried in order if primary fails
    - model: "qwen2.5:14b-instruct"
      url: "http://localhost:11434"

All models are swappable. Any Ollama embedding model, any OpenAI-compatible chat endpoint. See palinode.config.yaml.example for the full reference.

Embeddings without Ollama. llama.cpp (llama-server --embedding), vLLM, and LM Studio all expose the OpenAI-compatible /v1/embeddings shape; select it with dialect: openai (default ollama, so existing setups are unchanged). Retry, circuit breaker, and per-input error handling are identical to the Ollama path. The Ollama tag bge-m3 is not a llama-server model name — point llama-server at a BGE-M3 GGUF instead:

embeddings:
  primary:
    dialect: openai
    url: "http://localhost:8080"   # a trailing /v1 is fine too
    model: "bge-m3"                # llama-server ignores it; vLLM / LM Studio match it
    dimensions: 1024

When exposing the API beyond loopback (PALINODE_API_HOST other than 127.0.0.1), set PALINODE_API_TOKEN — the server refuses to start unauthenticated on a non-loopback bind unless you opt out explicitly with PALINODE_API_ALLOW_UNAUTH=1. See SECURITY.md for the bearer-token auth model and the bind gate.

API Reference

Design Principles

1. Files are truth. Not databases, not vector stores. Markdown files that humans can read, edit, and version with git.

2. Typed, not flat. People, projects, decisions, insights — each has structure. This enables reliable retrieval and consolidation.

3. Consolidation, not accumulation. 100 sessions should produce 20 well-maintained files, not 100 unread dumps.

4. Invisible when working. The human talks to their agent. Palinode works behind the scenes.

5. Graceful degradation. Vector index down? Read files directly. Embedding service down? Grep. Machine off? It's a git repo, clone it anywhere.

6. Zero taxonomy burden. The system classifies. The human reviews. If the human has to maintain a taxonomy, the system dies.

What's Unique

• Your data, your files — No accounts, no cloud dependency, no vendor lock-in. Your memory is markdown files in a directory you control. Export is cp. Backup is git push. Whatever happens to any tool in this ecosystem, your data is plain text on your filesystem.
• Cross-IDE memory — Your memory lives in one place. Connect from Claude Code, Cursor, Windsurf, Zed, or any MCP-compatible editor. Switch IDEs without losing context.
• Git operations as agent tools — diff, blame, rollback, push exposed via MCP. No other system makes git ops callable by the agent.
• Operation-based compaction — Structured operations are schema-checked and applied as reviewable git commits.
• Per-fact addressability — <!-- fact:slug --> IDs inline in markdown, invisible in rendering, preserved by git, targetable by compaction.
• 4-phase injection — Core (always) + Topic (per-turn search) + Associative (entity graph) + Triggered (prospective recall).
• Multi-transport MCP — stdio for local, Streamable HTTP for remote. One server, any IDE on any machine.
• If everything crashes, cat still works.

Measured, not asserted: docs/BENCHMARKS.md has LongMemEval results with methodology, cost, and the losses.

Acknowledgments

Palinode builds on ideas from Karpathy's LLM Knowledge Bases, Letta (tiered memory), and LangMem (typed schemas + background consolidation). See docs/ACKNOWLEDGMENTS.md for the full list.

See also the epistemic integrity discussion in the Karpathy gist thread — particularly the problem of LLM wikis that "synthesise without citing, drift from sources without knowing it, and present false certainty where disagreement exists." Git-based provenance is Palinode's answer to that problem.

If you know of prior art we missed, please open an issue.

License

MIT — Privacy Policy

Built by Paul Kyle with help from AI agents who use Palinode to remember building Palinode.