MGI-Mind

English | Русский | 中文

Latest release: v2.8.2 · CHANGELOG · Discussions · Issues · Contributing

<p align="center"> <a href="https://mcprush.com/madgodinc/mgi-mind-mcp"><img width="380" height="200" src="https://mcprush.com/badge/mgi-mind-mcp/card.svg" alt="MGI-Mind MCP server" loading="lazy" /></a> </p>

Local long-term memory for AI assistants. One Rust binary, a local Qdrant vector database, local ONNX models. Speaks MCP, so Claude Code and other assistants read and write memory on their own. Also a normal CLI.

<p align="center"> <img src="docs/demo.gif" alt="Store a fact, then recall it by meaning: a query with no shared keywords still finds it" width="900"> <br> <em>Store a fact in plain words, recall it with a query that shares none of them. Semantic search matches intent. Your assistant runs the same <code>mind_search</code> on its own over MCP. Regenerate with <code>vhs docs/demo.tape</code>.</em> </p>

The same exchange inside an assistant:

You:  what was the deploy server again?

Assistant (calls mind_search "deploy server"):
  -> "Deploy server is 10.0.0.5:8080, SSH as deploy@, key in vault"  (source: infra.md)

You:  right, thanks

Nothing leaves the box. Embeddings, search, reranking and the vault are all local. No cloud account, no API key, no telemetry.

---

Contents

• What it is
  • The validity model
• Why bother
• Quick start
• Using it
• How it works
• Command reference
• Configuration
• Languages and the reranker
• Changing the embedding model
• Troubleshooting
• Security
• Status and audit
• Project layout
• License

What it is

MGI-Mind sits between you and your assistant. The assistant writes short notes ("memories") and facts as a conversation goes, and pulls the relevant ones back when they matter.

Retrieval, not just storage:

• Hybrid search. Each memory is stored as two vectors: a dense vector (multilingual-e5-base, 768 dims) for meaning, and a sparse term-frequency vector (TF-IDF, BM25-style) for exact words. A query runs both arms and fuses them with Reciprocal Rank Fusion. "The server box" finds "deploy host" through the dense arm; fossilize_replay finds the one note that contains that exact token through the sparse arm.
• Cross-encoder reranking. Fused top candidates are re-scored by bge-reranker-base, which reads the query and each passage together and is more accurate than comparing vectors. On by default and English-tuned; see Languages and the reranker for the trade-off on other languages.
• One warm process. mgimind mcp is the MCP server itself: it runs for the whole session, models load once and stay warm in memory, so a lookup costs milliseconds instead of reloading on every call.

Around retrieval there's a knowledge graph for structured facts, per-agent session logs for cross-session continuity, and an encrypted terminal-only vault for secrets.

The validity model

Hybrid search is table stakes; most memory tools have some version of it. The part that is harder to find is what keeps the store from ossifying as it fills with old, contradictory, or self-reinforcing beliefs. These run mostly on their own:

• Duel rule. When a new fact contradicts an existing one on the same subject and predicate, the second write resolves against the first instead of piling up a second "truth". An entrenched fact (many dependants, confirmations, age) holds; a strong fresh fact flips it and dampens the loser to a hidden stale status; a borderline one is marked contested or diverted to quarantine. Confirmations accrue live: re-asserting the same triple raises its own confirmations_count, so a fact confirmed a few times resists a single fresh contradiction (contested, not flipped) while an unconfirmed young fact still flips as easily as before. dependants still only comes from the migrate-v14 dependants --apply backfill - a cheap live producer (it would mean walking the graph on every write) is not implemented. Nothing is deleted, so the audit log keeps the loser; a re-proposed stale loser gets a fair new duel against whatever is current, keeping its own original created_at and confirmation history rather than starting over. Automatic, inside a normal mind_fact add.
• Doubt window. An entrenched fact has to keep re-justifying itself. Every fact query measures how far the current context has drifted from the context the fact was written in; after five consecutive drifted retrievals the fact stops outranking facts that still hold up, and any single in-context retrieval resets the count. Nothing is hidden or deleted. Drift is measured on vectors with the corpus mean removed, because raw cosine cannot separate topics in an anisotropic embedding space (see mgimind calibrate below). The current context is the running process's own recent queries, so this applies inside a warm mgimind mcp or serve-http session; a one-shot CLI command has no context yet and no fact is penalised for that. This ships in shadow: drift is measured and recorded, and no fact changes, until you set doubt_drift_threshold from your own store's numbers and switch doubt_drift_mode to "enforce". A background pass separately re-tests entrenched facts that have gone quiet, under hard guarantees: never during a tool call, a per-tick cap, and a load-aware cadence.
• Inheritance discount. mind_context's briefing (what a session sees at start) marks every fact it surfaces as inherited for that process. If a duel challenger is a re-proposal of one of those inherited facts - the agent echoing back what it was told rather than independently reconfirming it live - that challenge counts at half weight. The flag clears the moment the agent actually re-asserts the fact live and it lands without a conflict. Single-source decay (repeated confirmation from the same source counting for less each time) has the formula (diversity_weighted_count) but no live caller yet - not implemented, no "cannot co-confirm" guarantee today.
• Two time axes on facts. Record-time (created_at/valid_until: when the store learned/retired a value) is bi-temporal with the superseded chain: a predicate can be registered (mind_predicate) as single-valued, temporal-single (one current value, but the previous ones are kept and queryable by date via mind_history), or multi-valued, and a superseded value is hidden from default ranking, not erased. An optional second axis, world-time (valid_from, settable on add), records when a fact became true in the world instead of when the store heard about it; mind_fact action=query with as_of + world_time=true resolves point-in-time queries against it, falling back to record-time per fact when it has no valid_from. Existing facts need no migration - they simply have no valid_from and behave exactly as before.
• Explicit invalidate is final. mind_fact action=invalidate is a one-way door: re-adding the same triple afterward refuses (verdict invalidated) instead of silently reviving it. mind_fact action=restore is the only way back.
• Typed outcome signals. mind_outcome records that a remembered fix actually worked (test_passed, code_compiled, user_confirmed, cited_by). A real success raises a memory's weight and marks a procedure verified; a failure pulls the weight down rather than being ignored.

These are research-shaped mechanisms, and their tuning constants are still being calibrated (the TODO(phase-4-calibration) markers in duel.rs / doubt.rs are honest). The shape is the point: a memory that argues with itself and demotes what stops holding up, instead of accreting forever.

mgimind calibrate measures whether that shape behaves as intended. It runs a corpus of realistic conflict situations through the live duel formulas and reports how many land on the outcome a person would expect: a fresh unsupported claim cannot overturn an entrenched belief, a CI signal can, repetition alone coexists rather than overwriting. Today 14 of 15 scenarios match intent; the one that does not is printed with its reason, not hidden, and the suite runs in CI. It also reports the doubt window's drift distribution, and that report exists because of what it found. The drift threshold was 0.4, picked in 384-dim MiniLM space and left in place when the default model became 768-dim multilingual-e5. Measured on a live store, 400 memories and all 79800 pairs, raw cosine ran from 0.666 to 0.972: the threshold could not fire once. The corpus mean vector had norm 0.884, so almost all of that cosine was a component every embedding shares. Centering fixes the measurement, and calibrate proposes a threshold from the drift your own store produces rather than from a constant that travelled between embedding spaces.

This measures the shape of the validity model, separate from retrieval recall (R@k, see BENCHMARKS.md). The two are different numbers and never belong in the same table.

Why bother

An assistant without memory asks for the same context every session and can't build on yesterday's work. The usual workaround is you keeping notes, tags, and folders, which the assistant still can't read by meaning.

The thing MGI-Mind does that Obsidian and Notion don't: the system decides what to write down, you don't. The MCP server reads what the assistant is doing in real time and routes facts, decisions, and fixes into the store through a relevance gate. You don't file, you don't tag, you don't decide what's worth saving. Low-signal candidates land in a quarantine layer (recoverable on re-assertion) instead of polluting retrieval.

How it compares to the obvious alternatives:

• Plain notes (Obsidian, Notion). Strong as your personal notebook, but the assistant can't search them by meaning, and every keystroke is equally important to a folder of .md files.
• Bare vector database. Semantic search, but no exact-term matching, no reranking, no dedup, no sessions, no facts, no secrets handling, no relevance gate, no procedural memory. You assemble all of that.
• Hosted memory API. Your data on someone else's servers. Closes the box on inspection, dedup behaviour, and what the relevance gate is actually doing.

MGI-Mind is the assembled local version: hybrid + reranked retrieval, the relevance gate, dedup, facts, sessions, procedural memory ("error → fix" playbooks), and a terminal-only vault, all behind one binary you run yourself.

A note on evaluating it. BENCHMARKS.md reports retrieval recall (R@k on LongMemEval-S, zero-API, no LLM judge): given a question, did the gold session land in the top-k. On the default CPU config that is R@5 ≈ 98%. This is not the same metric as the QA-accuracy numbers other memory tools publish (they run an answerer and a judge over their retrieval), so do not put it in the same table. It measures whether the evidence was retrievable, not whether an LLM then answered correctly. mgi-mind has no answering step, so retrieval recall is the number it actually owns.

The honest gap: the validity model (duel rule, doubt window, supersession) is the differentiator, and retrieval recall does not test it. Two things test it instead. mgimind calibrate checks behavioral shape aga