# Attestor A self-hosted memory service for agent teams: roles, write and forget permissions, version history, audit. Built for the case where several agents write to one store. - MCP server · `type=mcp` · Claude Code, Codex, Cursor, GitHub Copilot, Windsurf (Devin Desktop), Cline, Zed, Gemini CLI · ★ 14 - Author: bolnet - License: MIT · policy=link-only · description only; the file stays with the author - Source: https://github.com/bolnet/attestor - Updated: 2026-09-18 - HTML: https://affpapa.org/en/nevibecode/mcp/attestor ## What it does Attestor keeps each memory once in Postgres and derives vectors and a graph into Pinecone and Neo4j. Reads and writes pass through an `AgentContext` carrying the agent's identity, role, namespace and budget. Six roles exist: `ORCHESTRATOR` reads, writes and forgets; `PLANNER`, `EXECUTOR` and `RESEARCHER` read and write; `REVIEWER` and `MONITOR` only read. The `read_only=True` flag strips write and forget whatever the role, and a violation raises `PermissionError` before any backend is reached. Ranking on the read path is deterministic: vector, BM25, RRF, graph, MMR, then budget fit, with no language model in scoring. The same query returns the same order, which makes it unit-testable. An optional reranker (`bge`, `cohere` or `llm`) stays off unless you turn it on. Every fact carries valid time and transaction time, so `recall(as_of=...)` reconstructs what the mesh believed at an earlier point. Contradictions supersede older facts instead of deleting them. One caveat: the Pinecone lane ignores `as_of` and logs a warning, so a historical view through vector retrieval is not guaranteed. It ships as a Python library, a REST sidecar and an MCP server over one API. Role enforcement is an in-process guarantee; the REST and MCP surfaces hit the store directly and lean on their own auth. Install with `pip install attestor`; the MCP launch uses `uvx`. Version 4.1.11 is a greenfield rebuild, and there is no migration path from v3. ## Requirements - Needs: Python and uv (commands run via uvx) - Environment variables: `ATTESTOR_PATH`, `ATTESTOR_DISABLE_LOCAL_EMBED`, `OPENAI_API_KEY`, `OPENROUTER_API_KEY` — keys and tokens: ask the human for the values, never commit them or put them in shared files. ## Install ### Claude Code (native platform) **Platform CLI** ```bash claude mcp add --env ATTESTOR_PATH= --env ATTESTOR_DISABLE_LOCAL_EMBED= --env OPENAI_API_KEY= --env OPENROUTER_API_KEY= --transport stdio attestor -- uvx attestor ``` Replace <…> with your own values. uvx runs the command named after the package; if the README names another one, use uvx --from attestor ### Codex **Platform CLI** ```bash codex mcp add attestor --env ATTESTOR_PATH= --env ATTESTOR_DISABLE_LOCAL_EMBED= --env OPENAI_API_KEY= --env OPENROUTER_API_KEY= -- uvx attestor ``` Replace <…> with your own values. uvx runs the command named after the package; if the README names another one, use uvx --from attestor ### Cursor **Config file** File `.cursor/mcp.json`: ```json { "mcpServers": { "attestor": { "command": "uvx", "args": [ "attestor" ], "env": { "ATTESTOR_PATH": "", "ATTESTOR_DISABLE_LOCAL_EMBED": "", "OPENAI_API_KEY": "", "OPENROUTER_API_KEY": "" } } } } ``` Replace <…> with your own values. uvx runs the command named after the package; if the README names another one, use uvx --from attestor ### GitHub Copilot **Config file** File `.vscode/mcp.json`: ```json { "servers": { "attestor": { "type": "stdio", "command": "uvx", "args": [ "attestor" ], "env": { "ATTESTOR_PATH": "", "ATTESTOR_DISABLE_LOCAL_EMBED": "", "OPENAI_API_KEY": "", "OPENROUTER_API_KEY": "" } } } } ``` Replace <…> with your own values. uvx runs the command named after the package; if the README names another one, use uvx --from attestor ### Gemini CLI **Platform CLI** ```bash gemini mcp add -e ATTESTOR_PATH= -e ATTESTOR_DISABLE_LOCAL_EMBED= -e OPENAI_API_KEY= -e OPENROUTER_API_KEY= attestor uvx attestor ``` Replace <…> with your own values. uvx runs the command named after the package; if the README names another one, use uvx --from attestor ## When to use it - decide which agents may write to a shared memory - show what an answer was built from months later - reconstruct memory state as of a past date - delete a user's data with an audit trail - keep tenants separated inside one store ## Example prompts - Add the auth schema decision to memory as agent `reviewer-1` - Show what the agent team believed last Friday - Forget this user's data and show me the audit entry ## Similar - [supabase-audit-rls](https://affpapa.org/en/nevibecode/skills/supabase-audit-rls.md) — skill · 1100 installs - [supabase-security-audit](https://affpapa.org/en/nevibecode/commands/supabase-security-audit.md) — command · 1020 installs - [supabase-audit-tables-read](https://affpapa.org/en/nevibecode/skills/supabase-audit-tables-read.md) — skill · 487 installs - [codex-memory-curator](https://affpapa.org/en/nevibecode/skills/codex-memory-curator.md) — skill · 2700 installs - [recall](https://affpapa.org/en/nevibecode/skills/parcadei-recall.md) — skill · 493 installs - [YourMemory](https://affpapa.org/en/nevibecode/mcp/yourmemory.md) — MCP server · ★ 270 --- NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md