接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令？ 兩種安裝方式的差異
來自發布者
Local, deterministic memory for coding agents.
mind stores durable project facts in a weighted concept graph, recalls them
through lexical ranking plus spreading activation, tracks where every fact
came from, ages unused salience, and consolidates recurring themes through a
deterministic dream cycle.
Persistent memory for AI coding agents: automatic capture, explainable recall, knowledge consolidation, privacy controls, and portable offline storage. One Python file, zero dependencies.
The default runtime is:
mind.py file.The development source is split by domain under src/mind/. A deterministic
builder reconstructs the single-file artifact byte-for-byte.
7.0.1.dev0 (preview).7.0.0; pinned mind.py SHA-256 ae2fc389b3b09c93cb432ab55b71063d98b400da6b18d6bc178322bc8f3fcf69.182322c812686d5885423f90943a66898236ec04a1ed84ce5b52feb9a4364aee.7.0.1.dev0 is the current development preview. The production-ready release
is 7.0.0, which contains the lifecycle, protocol-server, typed-memory,
privacy, automatic-capture, and modular-source features documented below.
curl -fsSLO https://raw.githubusercontent.com/Da7-Tech/mind/v7.0.0/mind.py
python3 -c "import hashlib; p=open('mind.py','rb').read(); assert hashlib.sha256(p).hexdigest() == 'ae2fc389b3b09c93cb432ab55b71063d98b400da6b18d6bc178322bc8f3fcf69'"
python3 mind.py init
git clone https://github.com/Da7-Tech/mind.git
cd mind
python3 tools/build_single.py --check
python3 mind.py init
On stock Windows, exported commands use py -3 mind.py. A Windows CI field
test converts the artifact to CRLF, runs the exported invocation verbatim, and
checks the resulting project.
Facts can be semantic, episodic, procedural, or decisions. They also carry
scope, authority, source trust, sensitivity, expiration, pinning, and optional
entity/attr slots for contradiction detection.
mind init exports a guard-marked standing-order block into AGENTS.md,
CLAUDE.md, and GEMINI.md. Existing content outside the generated block is
preserved byte-for-byte. Existing Cursor, Windsurf, Cline, and Roo rule files
are adopted when present.
Agents use:
python3 mind.py capture "durable project fact"
Automatic capture:
Dream scheduling is independent from the telemetry log. A bounded scheduler uses a lease, pending count, and recovery rules. Oversized telemetry resets safely without disabling future maintenance.
Host integrations can consume machine-readable recipes:
python3 mind.py integrations --json
python3 mind.py context --json
The recipes cover session start, durable capture, pre-compaction batch flush, session end, an optional scheduled backstop, and the protocol server.
init
--help
--version [--verbose]
remember "text"
remember --user "text"
remember --json
remember --batch
capture "text" [--trust LEVEL]
pending
approve ID
reject ID
context [--json]
suggest-user [--json]
integrations [--json]
recall "question" [--at DATE] [--explain]
confirm ID [...]
correct "old hint" "new fact"
link "a" "b" [relation]
forget ID [--reason TEXT]
unlink A B
redact ID --reason TEXT
purge ID|--match TEXT --all-traces [--confirm]
why ID
entity "term"
dream [--dry-run]
backup [label]
checkpoint [label]
restore NAME [--confirm]
compact [--dry-run] [--keep-journal-days N]
merge BASE OURS THEIRS [--output PATH] [--graph-out PATH]
doctor [--bench] [--json]
growth [--days N] [--json]
export
status
mcp
Use recall -- "-query beginning with a dash" for a dash-leading query.
Offline recall combines:
recall --explain prints direct, spread, fused, semantic, and final scores,
plus backend identity, process calls, latency, and fallback reason. Recall is a
pure read; useful hits change durability only after explicit confirm.
Directional relations such as depends-on, owned-by, and deployed-to
store truthful reverse labels while retaining bidirectional traversal.
The default remains offline. Two explicit optional protocols exist:
MIND_EMBED_CMD: one versioned batch process per ranking;MIND_EMBED_SERVER: a persistent length-framed process with handshake,
model revision, and dimension identity.Both paths have bounded output, vector dimension, cache bytes, and total ranking deadlines. A partial failure falls back the entire ranking to the offline metric; similarity spaces are never mixed.
Reference server:
export MIND_EMBED_SERVER='python3 contrib/concept_embed_server.py'
python3 mind.py recall "where are backup copies kept" --explain
The process receives the query and candidate memory text. Only configure a program you trust. The tool does not enforce that program's network isolation.
python3 mind.py mcp
The same file serves newline-delimited JSON-RPC over standard input/output. It supports initialization, ping, tool listing/calls, cancellation notifications, clean EOF shutdown, and seventeen memory, diagnostic, and privacy tools. Standard output contains protocol JSON only.
Minimal lifecycle:
initialize
notifications/initialized
tools/list
tools/call
notifications/cancelled
status separates current journal bytes from segment count and segment bytes.
compact rotates an oversized active archive, segments a current journal when
it exceeds budget or is wholly older than the requested retention horizon, and
collects stale temporary files.
Backups are plain files with a SHA-256 manifest:
python3 mind.py backup before-upgrade
python3 mind.py restore BACKUP_NAME
python3 mind.py restore BACKUP_NAME --confirm
A confirmed restore creates a pre-restore checkpoint first. Privacy rewrites also refresh backup manifests so a remediated backup remains verifiable. Restore writes an exact file plan, removes later managed files, and resumes an interrupted plan before normal memory loading.
forget removes a fact from retrieval but keeps an auditable tombstone.unlink removes a relation without deleting its endpoints.redact replaces payloads with a digest and reason across managed stores.purge inventories first; --confirm irreversibly removes payload and node
identifiers from graph, journals, archives, dreams, cortex, exports, queues,
receipts, and backups.Redaction and purge use a crash-resumable outbox. Exact-byte tests search every managed artifact after completion. Secrets should still never be stored: remediation is a last resort, not a secret manager.
Journal format v2 adds UTC epoch-nanosecond time and stable event IDs. The three-way merge command deduplicates suffix events, orders them deterministically, and can replay the merged journal into a graph:
python3 mind.py merge BASE OURS THEIRS --output MERGED --graph-out GRAPH
Example merge driver:
[merge "mind-journal"]
name = deterministic mind journal merge
driver = python3 mind.py merge %O %A %B --output %A
.mind/journal.jsonl merge=mind-journal
On Windows, replace python3 with py -3.
python3 mind.py doctor --bench
python3 mind.py growth --days 30
python3 mind.py suggest-user
doctor checks storage boundaries, recovery outboxes, scheduler leases,
duplicate export guards, BOM/CRLF handling, stale temporary files, backend
configuration, and clock anomalies. Its optional personal benchmark appends a
local recall history.
growth derives learned, confirmed, corrected, forgotten, dreamed, promoted,
and conflict counts from journal and dream truth. The latest bounded
consolidation receipt is visible in every generated ACTIVE.md.
Every public result is JSON tied to an immutable input, source identity, backend identity, and exact command.
On this subset, BM25 leads both evidence metrics. This benchmark does not measure graph traversal, temporal validity, contradiction handling, or lifecycle operations, so it does not establish overall product superiority.
LongMemEval values are evidence@1 / evidence@5 / answer-string@5.
Raw files live under bench/results/. The LongMemEval input is pinned by
revision and SHA-256 in bench/manifests/longmemeval.json.
LongMemEval memory records include metadata prefixes such as date, question ID, session ID, and role. These prefixes improve provenance and isolation but also add query-independent tokens; evidence metrics therefore use exact node labels, and answer-string metrics are reported separately.
Mutation analysis performs a green product-suite baseline preflight, stages every product-test dependency, mutates modular source and artifact consistently, preserves bounded diagnostics, and distinguishes killed, survived, timed out, compile error, and infrastructure error. The three self-referential public-evidence tests run after report generation because they validate the completed report itself; every report records this exclusion and the exact baseline test count.
python3 tools/build_single.py --check
python3 tools/claims.py check
python3 -m unittest discover -s tests -v
python3 bench/bench.py
python3 bench/multilang.py
python3 bench/discrim.py
python3 bench/slots.py
python3 bench/soak.py
python3 bench/fuzz.py --quick
python3 bench/autonomy.py --quick
Release-only gates also run the full five-year autonomy horizon, immutable LongMemEval subset, both mutation targets, privacy scan, and all nine CI cells. See the V7 verification record for the three-method gate, raw evidence, limitations, and remote completion requirements.
Read SECURITY.md before enabling an external semantic backend or storing sensitive project material. The default kernel does not use the network or child processes. Optional semantic modes execute a trusted local program and pass memory text across that process boundary.
MIT licensed.
Persistent memory for AI coding agents: automatic capture, explainable recall, knowledge consolidation, privacy controls, and portable offline storage. One Python file, zero dependencies.
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式，而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Mind 透過 stdio 在本機執行。
上次檢查時間：2026-10-27。 Persistent memory for AI coding agents: automatic capture, explainable recall, knowledge consolidation, privacy controls, and portable offline storage. One Python file, zero dependencies.
