Perseus Vault

Perseus Vault is a persistent memory store for AI agents: one Rust binary writing to one SQLite file, with no container, external database, or cloud service. Memory operations are exposed through a canonical Model Context Protocol (MCP) surface, so any MCP host reads and writes the same store, and stored entity bodies are encrypted at rest with AES-256-GCM while the search index and metadata stay in plaintext.

Retrieval is hybrid: BM25 keyword search over SQLite FTS5, dense embeddings, and reciprocal rank fusion across both. History is bi-temporal in the SQL:2011 sense — every fact carries a transaction time (what was believed at an instant) and a valid time (what was true in the world at an instant) — so retroactive corrections, future-dated facts, and out-of-order arrival are recorded without erasing the earlier record.

The server runs as MCP stdio, and one install-client command registers it with detected hosts, points them at a single shared database, installs session-lifecycle hooks for recall at session start and hygiene at session end, and appends memory usage rules to the agent instruction files. Adapters make it the memory backend for LangGraph as a BaseStore, CrewAI as an agent tool, and AutoGen as a Memory implementation. HTTP or SSE transport is opt-in and must be authenticated by the deployment.

The advertised tool surface is release- and profile-specific: the v2.23.2 --no-default-features snapshot publishes 175 canonical tools, and --profile lean narrows the advertised list to the core memory calls while hidden tools stay callable. Tools cover entity storage and time travel, search and retrieval-augmented generation, a typed link graph with community detection, an append-only journal, key-value state with time-to-live, lifecycle grooming through decay, pruning, archiving and consolidation, quality scoring and conflict detection, policy keystones, and file-based export and import.

Published measurements drive the shipped binary over MCP stdio. On the public LongMemEval _s split (500 questions, 23,867 sessions), session-level retrieval reaches:

Path	recall@1	recall@5	recall@10	MRR
keyword only (FTS5)	4.2%	19.2%	33.6%	0.1069
dense	75.8%	91.8%	96.0%	0.8296
hybrid (RRF)	83.2%	98.8%	99.8%	0.8949

Features

• Single-binary deployment: one executable and one file, with no Docker, Postgres, or cloud dependency
• Hybrid recall: BM25 full-text search, dense vectors, and reciprocal rank fusion, plus a dense-only semantic search shortcut
• Bi-temporal history: asof, validat, and bitemporal queries answer what was believed at a time and what was true at a time
• Encryption at rest: AES-256-GCM over entity bodies, with an owner-only key file created on first run
• MCP-native surface: canonical perseusvault* tools over stdio, with lean, agent, ops, and full advertisement profiles
• Client wiring: one command registers the server with detected hosts, shares one database, and installs session lifecycle hooks
• Framework adapters: LangGraph BaseStore, CrewAI tool, and AutoGen Memory implementations, each over an MCP stdio session
• Entity graph: typed links, traversal, community detection, community summaries, and global search across clusters
• Lifecycle grooming: decay scoring, pruning, archiving, expiry, consolidation, redaction, and physical erasure
• Workspaces and memory banks: a bank name maps deterministically to a workspace hash that scopes every read and write, with per-workspace authority manifests
• Journal and audit: append-only events with actor attribution, timeline queries, provenance tags, and a graph drift report
• Embeddings: a bundled local model, Ollama, or an OpenAI-compatible endpoint