Memory Vault

An open-source MCP server that gives Claude Code persistent memory by keeping project context in a local Postgres database instead of the model’s context window. Repo: github.com/mihaibuilds/memory-vault. Paged here from a MakeUseOf how-to (the source actually read; the repo is the canonical artifact).

The problem it targets

A coding agent’s working memory is the harness‘s finite context window. As it fills with file reads, tool output, and chat history, Claude Code auto-compacts — dropping older tool results and summarizing earlier turns. That is lossy: instructions and details “quietly get lost.” Memory Vault moves the durable context out of the window so it survives across sessions, /clear, and even a change of machine — the agent-memory layer made external and persistent.

How it works

• Store: Postgres + pgvector, doing hybrid search — semantic-similarity (vector) and keyword matching combined — so a saved memory is retrievable both ways.
• Transport: wired to the agent over the MCP. Configure .claude.json with the MCP server block pointing at the local Postgres instance (default credentials memory_vault/memory_vault). This is the spoke’s recurring “skills × MCP = agency” shape applied to memory — the store reaches the agent as an MCP tool.
• Deploy: git clone … && docker compose up -d — three commands. A web dashboard at localhost:8000 browses stored memories; docker compose exec app memory-vault status verifies the service. (Windows needs an asyncio fix for psycopg3.)
• Team angle: because the store is a shared database, multiple developers can read and write a collective project history — memory as shared infrastructure, not a per-developer file.

Where it sits

The article frames it against Claude Code’s built-in CLAUDE.md — a plain-text file auto-loaded each session (the AGENTS.md-style project-context convention). That works but consumes tokens every session; Memory Vault keeps context out of the window and fetches it on demand instead.

It is the second standalone memory-as-infrastructure product in the spoke after seekdb. Where seekdb is an ambitious MySQL-compatible DB with copy-on-write FORK/MERGE sandboxes, Memory Vault is the lightweight end: stock Postgres + pgvector, Docker-deployed, and delivered to the harness over MCP rather than queried as a database. Same layer, opposite weight class.

Provenance & caveat

Tier T3 — a MakeUseOf install-walkthrough (secondary how-to content), not the project’s own docs or an independent evaluation. Every claim here traces to that article; the tool author (mihaibuilds) is a thin GitHub handle, recorded inline rather than as a node. No benchmark of retrieval quality or the team-collaboration claim — recorded as published.

Related

agent-memory · seekdb · model-context-protocol · agents-md · durable-agents · openhuman · anthropic