One durable memory across supported coding agents.
Your Markdown vault is canonical. DuckDB is the replaceable retrieval cache.

Website · PyPI · Obsidian companion · Benchmarks

A cairn marks a trail for whoever comes next. agentcairn does that for coding agents: it captures durable context from the tools you use, stores it as inspectable Markdown with provenance, and recalls only the most relevant pieces when another agent needs them.

Proof you can inspect

The memory is not hidden behind an admin console or a hosted database. The separate agentcairn-obsidian companion reads the same Markdown files as the agents and exposes provenance, currency, importance, supersession, and related: links.

_{A real agentcairn vault in Obsidian. The list is a view over the files—not a second memory store.}

Dogfood snapshot · 2026-07-15. Across 417 local recalls, the maintainer's vault returned context about 262× smaller than loading the full vault each time—an estimated 136.6M tokens of full-vault context avoided in aggregate. Token counts use approximately four characters per token. This is not billed-token savings, and agentcairn sends no telemetry.

Install

The shortest path is a first-class plugin. It bundles the MCP server, the memory skill, and the host-specific ambient hooks—no separate agentcairn package install. The plugin launches through uvx, so install uv first if uvx --version is not already available.

Claude Code

claude plugin marketplace add ccf/agentcairn
claude plugin install agentcairn@agentcairn

Claude Code gets per-turn project-scoped recall, session/compaction capture, and the /agentcairn:recall, /agentcairn:remember, /agentcairn:memory, /agentcairn:savings, and /agentcairn:ingest commands.

Codex

codex plugin marketplace add ccf/agentcairn
codex plugin add agentcairn@agentcairn

Codex gets the bundled MCP tools and memory skill, live-verified SessionStart recall, and SessionEnd capture with cairn sweep as the out-of-band backstop.

Agent-assisted setup

Already use skills.sh or a find-skills workflow? Install the public setup assistant:

npx skills add ccf/agentcairn --skill agentcairn-setup -g

Then ask your agent: Use $agentcairn-setup to preview, install, and verify AgentCairn for this coding agent.

This installs setup guidance only—not the AgentCairn runtime, MCP server, plugin, or hooks. The assistant delegates those changes to AgentCairn's preview-first native installer and verifies the resulting integration. The Claude Code and Codex plugin commands above remain the shortest path.

The default vault is ~/agentcairn and is created on first use. A new empty vault has nothing useful to recall yet, so prove the whole loop explicitly:

You   → Remember this durable fact: staging deploys use blue-green.
Agent → written and indexed
You   → Recall the staging deploy strategy.
Agent → staging deploys use blue-green.  ↳ <memory permalink>

remember writes the Markdown note and index entry together, so immediate recall is part of the contract. The first local run may download and warm the configured embedding/reranking models.

The contract

How it works

• Capture: host hooks improve immediacy; cairn sweep reads supported transcript stores out-of-band as the durable backstop. AgentCairn redacts recognized credentials, deduplicates, importance-gates, and distills before its automated plaintext writes.
• Reconcile: the first read transactionally brings the vault-scoped index in sync with Markdown. A failed rebuild preserves the last good cache and the durable files remain untouched.
• Recall: BM25 and semantic vectors are fused with Reciprocal Rank Fusion, then optionally reranked. Model/provider failures visibly fall back to BM25 with diagnostics instead of returning incompatible vectors.
• Remember: the MCP tool atomically writes a Markdown note and updates the index under one writer lock, making a successful save immediately recallable.

Designed for trust

• Local by default. FastEmbed runs locally, the MCP server uses stdio, there is no required daemon or external database, and there is no telemetry.
• Plain boundaries. The synced vault contains Markdown; by default, the rebuildable .duckdb index stays outside it. Vault symlinks that escape the configured root are rejected.
• Time-aware corrections. valid_from, valid_until, and superseded_by keep old evidence visible while making current facts rank first.
• Deterministic graph. [[wikilinks]] and optional cairn link neighbors create an Obsidian-native graph without asking an LLM to invent entities.
• Project-aware recall. The current project is boosted by default; cross-project results remain available and are labeled. Automatic recall is project-scoped unless you explicitly opt into all projects.

Agents supported

Every host resolves the same configured vault. cairn install previews detected hosts without writing. MCP configuration writes are backup-first and preserve unrelated servers; plugin-host installs delegate to the host's own CLI.

Codex SessionStart was verified live end-to-end with agentcairn 0.24.2 / plugin 0.1.2. The installed SessionEnd command dispatch and detached sweep pass exact handler probes; cairn sweep remains the out-of-band capture backstop. See the OpenCode integration and Hermes integration for their native lifecycle details.

Using it directly

The plugin is the easiest route, but agentcairn is also a standalone CLI and on-demand MCP server. Standalone installs require Python 3.11+.

uv tool install agentcairn

cairn init ~/agentcairn
cairn sweep --vault ~/agentcairn
cairn recall "how did we fix the auth bug?" --vault ~/agentcairn
cairn doctor --vault ~/agentcairn

Bring Claude Code's memory with you

Claude Code's auto-memory can seed the shared vault without changing its source files. The command previews only the current repository by default; add --apply to write the redacted notes and refresh the index.

cairn import claude-memory                         # preview; writes nothing
cairn import claude-memory --apply                 # import this repository
cairn import claude-memory --project ../other --apply

The one-way import reads MEMORY.md and its topic Markdown files—never CLAUDE.md or .claude/rules/. Imported notes retain Claude Code, project, and source-file provenance. When a source changes, the prior version remains inspectable but is superseded; when one disappears, its imported version expires. A small .agentcairn/native-memory/ registry preserves that lifecycle without indexing source content twice. Use --source <dir> for a custom, managed, or session-overridden Claude memory directory, or --no-reindex when batching imports.

Prefer an ephemeral process:

uvx agentcairn                             # MCP server
uvx --from agentcairn cairn recall "..."  # CLI; plain `uvx cairn` is a different package

cairn schedule install --vault ~/agentcairn  # launchd on macOS / user crontab on Linux
cairn schedule status
cairn link --vault ~/agentcairn              # write deterministic related: neighbors
cairn reindex ~/agentcairn                   # rebuild the disposable cache
cairn savings                                # local context-efficiency estimate
cairn index-status --vault ~/agentcairn

cairn config --init
cairn config

auto_recall = true
auto_recall_k = 3
auto_recall_scope = "project"  # use "all" only as an explicit cross-project opt-in

Benchmarks measured

The repository ships a revision-pinned, reproducible LongMemEval-S + LoCoMo harness. The default is local nomic-embed-text-v1.5 plus the cross-encoder reranker.

Context returned at the default k=10 is much smaller than the complete indexed history:

Read the numbers honestly:

• Retrieval recall is not QA accuracy. These tables compare controlled retrieval arms, not end-user answer quality or another product's leaderboard score.
• Token counts use an approximately four-characters-per-token heuristic. The reduction compares the indexed haystack with returned chunks; it is not billed cost savings.
• Graph boost is inert on these chat corpora because they contain no native [[wikilink]] graph. It is designed for real interlinked vaults.
• The optional QA judge uses Anthropic rather than the papers' GPT-4o setup, so those QA results are useful for relative ablations—not published-leaderboard comparisons.

Full metrics, embedding sweeps, latency measurements, licenses, commands, and caveats live in benchmarks/README.md.

Privacy and limits

• The vault is plaintext by design, not encrypted storage. AgentCairn redacts recognized credential patterns before its automated body/title/tag writes; unknown patterns and hand edits remain your responsibility.
• Vault files are owner-only (0600/0700). Because the vault is plaintext and redaction is best-effort, the file mode is effectively its only access control. Shared-GID setups (e.g. two Docker containers on the same group but different UIDs) need group access, so vault_group_writable = true widens new vault notes and directories to 0660/0770. It is opt-in on purpose: on macOS every local user's primary group is staff, so a group-readable default would expose your memories to other accounts on the machine. The knob never widens anything outside the vault — the index, ledgers, lock files, and ~/.agentcairn/config.toml stay private.
• Cloud features are explicit egress. The default stays local. Opting into a cloud embedder or LLM judge sends the remaining redacted text to that provider.
• The project is beta. Standalone use requires Python 3.11+, and the first local model load can take time. The published retrieval evidence is strongest for conversational memory, not a universal code-search claim.
• Ambient behavior varies by host. The matrix above is intentional: Cursor and Antigravity rely on sweep capture; generic MCP hosts may expose tools without lifecycle hooks.
• Automation is platform-specific. Managed scheduling targets macOS launchd and Linux user crontab; use your own scheduler elsewhere.

Development

agentcairn uses uv exclusively for dependency management and tooling.

uv sync
uv run pre-commit install

uv run pytest
uv run ruff format .
uv run ruff check --fix .
uv run pre-commit run --all-files

Run the offline benchmark regression without API keys:

uv run pytest benchmarks/tests/

License

Apache License 2.0 — permissive, with an explicit patent grant. Copyright © 2026 Charles C. Figueiredo.