Every memory tool stores what happened. The interesting move here is the procedure record: version three of "deploy" arrives carrying the assumption that turned out false — the build container had enough memory — and a precondition that travels with it. An agent picking it up does not have to re-derive the two mistakes that produced it. The rest is a competent hosted memory service; this is the part that is hard to copy.
A hosted memory service reachable over MCP, plus Python and JavaScript SDKs and a CLI. What separates it from a fact store is the third memory type: procedures carry a version, and reporting a failure produces a new version with the violated assumption and a derived precondition attached, so an agent loading the latest one knows what to check first.
- 30 tools for memory management over MCP, from Claude Desktop, Cursor, Codex, Windsurf, Cline or a hosted Claude agent
- Semantic memory for facts and preferences, episodic memory for events and decisions with their outcomes, and procedural memory for workflows
- Procedures that evolve: report an outcome and a new version is produced carrying `violated_assumption`, `preconditions`, and running `success_count` / `fail_count`
- Automatic failure detection — describing a failed deploy in conversation links the episode to the procedure and evolves it without an explicit feedback call
- A synthesised answer with citations that link each claim back to the fact it came from, rather than a raw list of matches
- A cognitive profile generated from all memories in one call, ready to paste into a system prompt
- Retrieval across 23 languages: ask in Russian, Chinese, Spanish or Japanese and get an answer in the same language, cited back to facts stored in their original one
- Multi-user isolation on a single API key, so each `user_id` sees only its own memories and profile
- Claude Code hooks that load the profile at session start, recall on every prompt and save after responses — surviving `/clear` and auto-compaction
- History import from ChatGPT exports, Obsidian vaults or plain markdown, with secrets redacted on your machine before upload
A Mengram API key, in the `om-` form, free from mengram.io. `server.json` publishes a streamable-http remote at `https://mengram.io/mcp` that takes your key as a bearer token in an `Authorization` header; the local path is `pip install mengram-ai` (2.30.0) and running the `mengram` command with `server` and `--cloud`, with `MENGRAM_API_KEY` set. `mengram try` runs entirely on your machine with no account if you want to see what it would store first. Self-hosting against Ollama is supported but the extraction prompt is around 4,000 tokens, so the project asks for a model of 8B parameters or more with an 8K or larger context window. Licensed Apache 2.0.
One command plus a key — pip install mengram-ai, then supply credentials
