Local cognitive memory for MCP-compatible agents, shipped as a single ~25MB Rust binary with a 13-tool MCP server (consolidated from 34 tools in the v2.2.0 'Tool Consolidation' release — old names remain dispatchable as hidden back-compat aliases), an Axum HTTP/WebSocket server, and a SvelteKit 3D memory dashboard. Implements neuroscience-grounded mechanisms — FSRS-6 spaced repetition, prediction-error gating, synaptic tagging, spreading activation, dual-strength model, Retroactive Salience Backfill, and 'memory dreaming' consolidation — across ~30 stateful cognitive modules. 100% local.
- Storage
- SQLite + FTS5 (optional SQLCipher encryption) for entities and keyword search; USearch HNSW for vectors. Prediction-Error Gating decides whether new info is merged, superseded, or stored; FSRS-6 (21 parameters) governs natural decay so unused memories fade; an active-forgetting `suppress` tool applies reversible top-down inhibition without erasing. Default embeddings are Nomic Embed Text v1.5 (768→256d Matryoshka, now F32 rather than the earlier I8 quantization); Qwen3-0.6B reranker optional. A `source_sync` tool can also index external systems of record (GitHub Issues, Redmine) into the same store, keyed by content-hash for idempotent re-sync.
- Retrieval
- A 7-stage cognitive search pipeline: HyDE query expansion (template-based, not LLM generation — expands query into semantic variants then embeds with local Nomic model) + keyword + semantic (HNSW, RRF fusion replacing the earlier linear score combination) + optional reranking + temporal + competition + spreading activation. All inference runs locally; zero cloud calls at query time. 'Memory dreaming' replays recent memories to discover and persist new graph connections; `deep_reference`/`recall` runs a trust-scored reasoning pass with claim-vs-memory contradiction detection (`claim_contradicts_memory`). The flagship `backfill` tool implements Retroactive Salience Backfill (Cai et al. 2024, Nature): on a recorded failure it reaches backward to promote the earlier, causally-upstream memory that produced it, even when the two share no vocabulary.
- Self-host
- Self-host: trivial
- License
- AGPL-3.0
- Pricing
- Open-source AGPL-3.0, free to self-host (`npm install -g vestige-mcp-server`); first run downloads a ~130MB embedding model, then fully offline. No paid tier. · Free / OSS
- GitHub stars
- 603
- Last release
- 2026-07-26
- Last commit
- 2026-08-03
- First catalogued
- 2026-06-28
Strengths
- Cognitive-science memory model: FSRS-6 decay, prediction-error gating, spreading activation, synaptic tagging, and consolidation 'dreaming' — not a vector store with decay bolted on
- Single self-contained Rust binary, 100% local (SQLite+FTS5 + USearch HNSW, optional SQLCipher encryption), with a 13-tool MCP server (consolidated from 34) across 8+ IDEs
- Retroactive Salience Backfill (Cai et al. 2024, Nature) reaches backward from a recorded failure to the causally-upstream memory that caused it, plus claim-vs-memory contradiction detection on write/recall — a genuine capability jump beyond a decaying vector store
- Active forgetting via a reversible `suppress` tool (top-down inhibition, 24h labile window) plus a real-time 3D memory dashboard
Watch out
- AGPL-3.0 copyleft: self-host freely, but offering Vestige as a network service requires open-sourcing your modifications — confirm license fit before commercial embedding
- Neuroscience mechanisms and retrieval gains are vendor-described; the README's own 'Silent Rotation' benchmark (Vestige 20/23 vs 0/23 for RAG) is a self-authored, self-reported result, not an independent/shared eval — record via harvest-benchmarks with selfReported:true, not here
- Single-maintainer project (~590 stars); some platform builds (Intel Mac, Windows MSVC+CUDA) need manual ONNX/toolchain steps
Best for
- Developers wanting a fully-local, inspectable cognitive memory for coding agents that decays, consolidates, and forgets like a brain
How it integrates
Benchmark results
No sourced results yet.
Sources
- Monthly card re-verification — GitHub API repo metadata: 595 stars, license AGPL-3.0 confirmed, archived=false, last commit on the default branch 2026-07-26 (third-party)
- Vestige README (vendor)
- GitHub API repo metadata (stars, AGPL-3.0 license, v2.1.27 release) (third-party)
- Vestige README — confirms HyDE is template-based (no LLM generation), local Nomic embeddings, zero cloud calls; Qwen3 Reranker optional per architecture diagram (vendor)
- Vestige README — npm install -g vestige-mcp-server@latest; MCP stdio transport; Claude Code, Cursor, VS Code Copilot, Windsurf, JetBrains, Xcode 26.3+, OpenCode integration commands (vendor)
- npm registry — vestige-mcp-server v2.1.27; bins: vestige-mcp, vestige, vestige-restore (third-party)
- Vestige README (monthly re-verify) — confirms 13-tool MCP server (down from 34), 25MB binary, Retroactive Salience Backfill / backfill tool, claim_contradicts_memory contradiction detection, and the self-authored 'Silent Rotation' benchmark (vendor)
- Vestige CHANGELOG — v2.2.0 'MCP Tool Consolidation (34 → 13 advertised tools)' and Retroactive Salience Backfill; v2.2.1 Windows embeddings + backfill safety fixes (vendor)
- GitHub API license endpoint — confirms AGPL-3.0 (third-party)
Last verified 2026-08-01 · updated by refresh-framework-cards