Vestige

Local-first memory for developers and their AI coding agents.

Vestige remembers project decisions across sessions, retrieves focused context, keeps source references current, and helps you investigate failures using earlier evidence. The open-source core runs locally through MCP, with automatic memory writes and an embedded dashboard. No API key is required for local memory and embeddings.

Install · Why not RAG · Benchmark · Science · Tools · Dashboard · Docs

New in v3

• Resume with current code references and focused, expandable context.
• Save memory automatically; inspect receipts and undo supported lifecycle changes.
• Keep durable recurring intentions across sessions, including snooze and explicit event triggers.
• Project selected project decisions and conventions into your client’s Markdown rules file.
• Merge duplicates from the dashboard and see when first-run embeddings are warming up.
• Use progressive tool discovery and smaller save responses to reduce repeated context overhead.
• Run the included developer benchmarks against your own tasks and usage data.

_{A labeled fixture store, a real run. The incident is fictional, seeded into a Vestige store with a seven month backdated timeline. The engine is not. A SIGSEGV on startup in an arm64 container, and the version pin set 23 days earlier that shares zero words with the failure. Similarity ranked the pin fourth. Backfill reached back, ranked it first, persisted the causal edge, and sealed the receipt. It names the suspects. It never calls the verdict. Watch the full 58 second walk, then run vestige backfill --contrast on your own store.}

Agents re-learn the same lessons: they recommend a change you already tested and rejected, re-derive a fix that was already written down, and treat every session as if the last one never happened. Vestige is the memory layer that ends that. Any MCP-capable agent (Claude Code, Claude Desktop, Codex, Cursor, and others) writes memories as you work and retrieves them later, modeled on real cognitive science: redundant memories merge, contradicted ones are flagged, unused ones fade, and when a failure hits, Vestige reaches backward to the decision that set it up.

The cause never looks like the bug. That is the whole product.

Install

You need Node.js. No Docker, no signup, no compile step (prebuilt for macOS ARM + Intel, Linux x86_64 + arm64, Windows x86_64).

Android (Termux) builds from source today; see docs/INSTALL-TERMUX.md.

npm install -g vestige-mcp-server@latest

Connect it to your agent. Every MCP client understands this config:

{
  "mcpServers": {
    "vestige": { "command": "vestige-mcp" }
  }
}

Verify: vestige dashboard, then open http://localhost:3927/dashboard. First run downloads a 130MB embedding model and, in the background, a ~150MB reranker, once; after that Vestige is fully offline, forever. Full walkthrough: docs/GETTING-STARTED.md.

Why not just RAG?

RAG retrieves text that resembles the query. That is the right tool when the answer looks like the question, and the wrong tool when the cause of a problem looks nothing like the symptom: a config choice from three weeks ago, a library pin, an assumption nobody flagged as risky.

The backward reach implements Retroactive Salience Backfill (Zaki, Cai et al., Nature 2024, 637:145-155, DOI 10.1038/s41586-024-08168-4): when a memory turns out to matter, the salience of the earlier memories that led to it is raised, so the causal chain becomes retrievable even though the surface text never matched. Every backfill result ships with a receipt naming the exact evidence path; Vestige reports receipt-backed candidate causes, never an unverifiable verdict.

And the limitation on the left column is not marketing: DeepMind proved single-vector retrieval mathematically incapable of certain relevance patterns (arXiv:2508.21038, ICLR 2026).

The receipts: Silent Rotation

The claim is testable, and the test ships with all 246 agent transcripts it produced. Three coding agents fix one failing e2e test; the fix needs the currently live signing key id, randomized per trial from a 50-key keyring, present in no file the agents can read. It exists only in the memory layer. The dangerous outcome is converging on a planted decoy: tests pass, the merge is clean, production breaks.

On the verbatim queries the agents typed, the causal memory ranks 7th of 8 under both dense cosine and BM25 while the decoy ranks 1st. Reproduce the central measurement in two seconds, stdlib only:

git clone -b benchmark/silent-rotation --depth 1 https://github.com/samvallad33/vestige.git
cd vestige/benchmarks/silent-rotation
python3 tests/bm25_baseline.py results/runA-trial-1/corpus-export.json --no-dense

The caveats are published alongside the results, including the trials a plain cosine baseline ties and the trial Vestige loses.

The science

Every mechanism is a cited result, implemented in Rust, running locally. Full write-up: docs/SCIENCE.md.

The tools

Your agent calls these; you rarely do.

For checkout-specific code advice, evidence states, explicit re-anchoring and startup response budgets, see Code context evidence.

For the installed action inventory, previews, typed intentions and compatibility changes, see Tool contracts.

Project scoping, hygiene workflows, and making memory a standing habit for your agent: docs/MEMORY_HYGIENE.md · docs/AGENT-MEMORY-PROTOCOL.md · docs/CLAUDE-SETUP.md.

The dashboard

vestige dashboard