agentsmemory

A multi-tenant memory palace for AI agents — served as a remote MCP server, backed by Ollama and a swappable vector index (Qdrant for the SaaS, an embedded one for self-hosted).

agentsmemory is the Go SaaS rewrite of the original Python mempalace: a semantic, long-term memory store that humans and AI agents read from and write to. Where the Python tool was a single local user with no auth, this is built for teams: each agent connects to a network MCP endpoint with a bearer token, operates inside its team's isolated workspace, and can pull centralised, versioned skills the team keeps up to date.

Status: early skeleton. The tenancy, auth, skill registry, storage clients and MCP transport are wired and verified end-to-end, and the core memory loop (file a drawer → recall it semantically) now works end-to-end against Ollama + the vector store. Today the server exposes 41 MCP tools, the same 41 whether hosted or self-hosted — the WRITE/FILE + SEARCH/RECALL families, the agent diary, the am_mine pipeline (text → chunked drawers + closet index), hybrid search (vector + BM25 + closet boost), the navigable graph (hallways + tunnels + traverse), the temporal knowledge graph, the skill-registry CRUD, and wing admin. Only the two single-user-local tools (sync, hook_settings) are intentionally not ported. See the Roadmap.