Wayminder: durable memory without the transcript archive

AI coding agents are useful inside one session, but the lessons that matter most often arrive just before the context disappears: a repository convention, an operational decision, or the cause of a recurring failure. Wayminder gives MCP clients a small, self-hosted place to keep that durable knowledge.

Memory with boundaries

A memory belongs to an explicit scope such as global, personal, a repository, or a host. Recall always enforces the effective scope: asking about one repository cannot turn into a search across every project. Stored entries retain provenance and version history, while corrections supersede old knowledge and deletion remains soft.

The service deliberately stores conclusions rather than conversations. Clients are told to recall before assuming, treat recalled text as untrusted data, and remember only verified facts, decisions, preferences, and reusable procedures. Likely credentials are rejected before they reach storage.

Local retrieval, standard protocol

Wayminder combines PostgreSQL full-text search with pgvector similarity and rank fusion. A local Ollama model creates embeddings, so memory content does not need to leave the deployment. Semantic deduplication keeps repeated observations from quietly filling the store.

The interface is Streamable HTTP MCP, with tools to remember, recall, list, supersede, forget, and inspect status. Registered clients authenticate independently; the server derives provenance from the credential instead of trusting a caller-supplied identity. The reference Compose stack keeps PostgreSQL and Ollama off the host network and exposes only the authenticated service.

Source and documentation: github.com/kilo666mj/wayminder.