agentbrain

Local-first long-term memory for AI agents — a plain Markdown vault + a thin MCP server. 给 AI Agent 用的本地长期记忆：纯 Markdown 知识库 + 薄 MCP server。

三步上手 / Quick start (3 steps)

pip install mnemosyne-lite             # Python >= 3.10（PyPI 包名；命令仍是 agentbrain）
agentbrain init ~/agentbrain  # 建立记忆库（幂等，可重复执行）
agentbrain doctor             # 自检：一切正常会显示 "Everything looks healthy."

{
  "mcpServers": {
    "agentbrain": {
      "command": "agentbrain",
      "args": ["serve"],
      "env": { "AGENTBRAIN_VAULT": "~/agentbrain" }
    }
  }
}

AGENTBRAIN_VAULT 可省略（默认 ~/agentbrain）；vault 在别处时才需要写，路径按你的实际情况改。把上面 JSON 粘进任意 MCP 客户端（Claude Code / Codex / Cursor / DSH / Open WebUI…），重启客户端，完成。Agent 从此有了跨会话、跨工具的长期记忆。

以后接入新的 agent 不用你教：对它说一句「读 AGENTS.md 照做」即可——文件开头会把新来者引导到 ONBOARDING.md，它自己就能判断接入状态（已接 MCP / 只有 shell / 只能读文件）并完成配置或降级。

Paste that JSON (with your vault path) into any MCP client and restart it — done. Your agents now share one long-term memory.

中文详细说明 · English quickstart

Why agentbrain / 设计理念

• Plain Markdown, no lock-in — your memory is a folder of .md files. Open it in Obsidian, grep it, version it with Git. Remove agentbrain and the memory stays.
• Token-efficient by design — index-first retrieval: Index.md is the cheap first layer, BM25 (CJK-aware) only ranks candidates, and query output is compact by default (mode='index'); full text only on demand.
• Append-only for agents — agents may create lessons, never edit or delete them. Consolidation happens through proposals in _consolidations/ that a human approves, which keeps multi-agent writes conflict-free.
• Plug-and-play via MCP — one server, every client: Claude Code, Codex CLI, OpenCode, Cursor, DSH, Open WebUI, ...
• Secrets never enter the vault — credentials live in env/keyring; lessons reference ${ENV:VAR_NAME} placeholders only, resolved at runtime via shell. Since 0.4.1 this is enforced, not just a rule: memory_ingest scans for credential-shaped content (sk-/ghp_/AKIA/xox-/AIza keys, bearer tokens, private-key blocks, password= assignments) and refuses the write, telling the agent to use a placeholder instead. Placeholders and teaching examples (sk-xxx, YOUR_KEY) ingest fine. lint also scans existing lessons and reports SECRET findings — read-only, never rewrites files. Hand-written files are never touched.

Vault layout

agentbrain/                    # vault root (git-friendly, Obsidian-friendly)
├─ AGENTS.md                     # rules every agent reads at session start
├─ ONBOARDING.md                 # one-shot access setup for new agents
├─ Case-Learnings/
│  ├─ Index.md                   # auto-generated lesson index (retrieval layer 1)
│  ├─ log.md                     # append-only audit log
│  ├─ Learnings/                 # one lesson per file, YAML frontmatter
│  │  └─ case-001-lesson-01.md   # 文件名 = {case_id}-lesson-{NN}，自动生成
│  └─ _consolidations/           # merge/promotion proposals (human approval)
└─ Agent-Profile/
   ├─ Immutable/                 # owner preferences & environment (agent read-only)
   ├─ Mutable-Hints/             # soft preferences (agent read-only)
   └─ _suggestions/              # agent-suggested profile changes

中文快速上手

pip install mnemosyne-lite             # Python >= 3.10
agentbrain init ~/agentbrain           # 生成 vault 脚手架（幂等），自动开启 Git 快照
agentbrain doctor                      # 体检：vault/索引/锁/快照/log 一览
agentbrain ingest --case demo --lesson "部署前必须先跑迁移脚本" --tags 部署,运维
agentbrain query "部署 迁移"
agentbrain profile                     # 查看个人偏好（Immutable + Mutable-Hints）
agentbrain suggest --title "回复用中文" --change "偏好简洁的中文回复"    # 提交偏好建议
agentbrain lint                        # 体检：重复/过时/无标签/低置信度 → 生成整合提案
agentbrain apply lint-20260820-172206.md      # 人工审核后执行提案（自动归档）
agentbrain verify my-case-lesson-01    # 重新确认某条经验仍然有效（消除 lint 的 STALE）
agentbrain distill                     # 分析 log 中重复出现的模式 → 生成提升提案
agentbrain snapshot -m "手动备份"      # 手动提交快照（如用 Obsidian 手改文件后）
agentbrain rules --agent trae --write  # 把记忆纪律写进客户端规则文件（项目根目录运行）
agentbrain rules --agent agentsmd --write  # agents.md 开放标准，覆盖 OpenCode/Gemini CLI 等
agentbrain rules --agent claude --write --global  # 写进用户级全局规则，所有项目生效

升级到新版本后（pip install --upgrade mnemosyne-lite）：

agentbrain upgrade                    # 刷新 vault 里的 AGENTS.md / ONBOARDING.md 模板
agentbrain rules --agent trae --write # 刷新各项目规则文件里的纪律块（旧版会原地更新）
# 然后重启你的 agent 客户端 —— MCP server 是进程启动时加载代码的

说明：PyPI 包名为 mnemosyne-lite（agentbrain 在 PyPI 上与已有项目过于相似，无法注册）。 安装后的 CLI 命令与 Python 包名仍是 agentbrain，GitHub 仓库地址不变。

日常你只需要做三件事（频率都很低）：

事	命令	频率
想看库健不健康	agentbrain doctor	随意
记忆整理（清重复/过时）	agentbrain lint → 审核 → agentbrain apply <提案>	约一周一次
手改文件后备份	agentbrain snapshot	改完就跑

其余全自动：Agent 会话开始读偏好、任务前查经验、学到东西写入（每次写入自动 git 快照，可回滚）。

在 MCP 客户端里接入（以 Claude Code 为例）：

claude mcp add agentbrain -- agentbrain serve

通用 MCP JSON 配置（Cursor / Open WebUI 等）：

{
  "mcpServers": {
    "agentbrain": {
      "command": "agentbrain",
      "args": ["serve"],
      "env": { "AGENTBRAIN_VAULT": "~/agentbrain" }
    }
  }
}

AGENTBRAIN_VAULT 可省略（默认 ~/agentbrain）；vault 在别处时才需要写。

Vault 路径解析顺序：--vault 参数 > AGENTBRAIN_VAULT 环境变量 > ~/agentbrain。

注册 MCP 只让 agent 能调记忆工具；要让它每次会话主动查库，再把纪律写进客户端的规则文件（在项目根目录运行，幂等可重复）：

agentbrain rules --agent claude --write   # 支持 claude / codex / trae / cursor / agentsmd（项目级）
agentbrain rules --agent agentsmd --write  # agents.md 开放标准：OpenCode、Gemini CLI、Amp 等都遵守
agentbrain rules --agent claude --write --global  # 用户级全局（~/.claude/CLAUDE.md），一次配置所有项目生效

一条命令把「任务开始查库、中途遇到新问题再查、学到就自主入库（无需确认）、密钥不入库」写进 CLAUDE.md / AGENTS.md / .trae/rules/ / .cursor/rules/。TRAE 和 Cursor 的全局规则在各自设置界面里，不走文件；claude / codex 支持 --global。

Onboarding a new agent later needs no instructions from you: just tell it "read AGENTS.md" — the file routes first-timers to ONBOARDING.md, where they detect their own access mode (MCP tools / shell / file-only) and wire themselves up or fall back accordingly.

English quickstart

pip install mnemosyne-lite             # Python >= 3.10
agentbrain init ~/agentbrain           # scaffold the vault (idempotent), enables git snapshots
agentbrain doctor                      # health check: vault, index, lock, snapshots, log
agentbrain ingest --case demo --lesson "Always run migrations before deploy" --tags deploy,ops
agentbrain query "deploy migrations"
agentbrain profile                     # print the owner profile
agentbrain suggest --title "Short replies" --change "Keep answers under 3 sentences."
agentbrain lint                        # health check → consolidation proposals
agentbrain apply lint-20260820-172206.md      # execute an approved proposal (archives it)
agentbrain verify my-case-lesson-01    # re-confirm a lesson is still valid (clears STALE)
agentbrain distill                     # recurring-pattern analysis → promotion proposals
agentbrain snapshot -m "manual backup" # commit a snapshot (e.g. after hand-edits)
agentbrain rules --agent claude --write # install memory discipline into the client's rule file
agentbrain serve                       # start the MCP server on stdio

After upgrading the package (pip install --upgrade mnemosyne-lite):

agentbrain upgrade                     # refresh AGENTS.md / ONBOARDING.md templates in the vault
agentbrain rules --agent claude --write # refresh the discipline block in project rule files
# then restart your agent clients — MCP servers load code at process start

Note: the PyPI distribution name is mnemosyne-lite (agentbrain was rejected as too similar to an existing PyPI project); the installed CLI command and the Python import name remain agentbrain.

Codex CLI (~/.codex/config.toml):

[mcp_servers.agentbrain]
command = "agentbrain"
args = ["serve"]

Registering MCP makes the tools available; making the client actually query at task start takes one more line — write the discipline block into the project's rule file (run in the project root, idempotent):

agentbrain rules --agent claude --write   # claude / codex / trae / cursor / agentsmd (per-project)
agentbrain rules --agent claude --write --global  # user-level (~/.claude/CLAUDE.md), all projects

It installs a short "agentbrain memory discipline" section into CLAUDE.md, AGENTS.md, .trae/rules/ or .cursor/rules/: query at task start, re-query on new subtasks/errors, ingest autonomously when something is learned, no secrets ever. The agentsmd target writes the agents.md open standard file, followed by OpenCode, Gemini CLI, Amp and other standard-compliant tools. TRAE and Cursor keep their global rules in their settings UIs; claude and codex support --global.

MCP tools

Tool	Purpose
memory_query(query, top_k=5, mode="index", tag=None)	Search lessons. mode='index' returns compact hits (id, summary, tags, path, gist); mode='full' adds full text; tag narrows results to one tag.
memory_ingest(case_id, lesson, tags, confidence=0.8, source_summary=None)	Save a new lesson (facts + scenario + fix, ≤ 30 lines). Creates a file, updates Index.md and log.md.
memory_lint(scope="all")	Health check: duplicates, stale, expired, untagged, low-confidence. Writes a merge proposal to _consolidations/.
memory_distill(window_days=30, min_repeat=3)	Finds cases/tags ingested ≥ N times in the window and writes a promotion proposal.
memory_profile()	Returns the owner profile (hard rules + soft preferences). Read-only; agents call it once per session to tailor behavior.
memory_suggest(title, change)	Proposes a profile change into Agent-Profile/_suggestions/ for the owner to review — agents never edit the profile itself.

MCP resources

URI	Content
agentbrain://rules	AGENTS.md — vault rules for every agent
agentbrain://index	Case-Learnings/Index.md — retrieval layer 1
agentbrain://profile	merged owner profile (read-only)

Agents are expected to follow AGENTS.md in the vault root: read the profile at session start, query at task start, ingest on learnings, never edit existing lessons, never write secrets into the vault. Consolidation proposals carry machine-readable directive blocks (```agentbrain); only the owner executes them via agentbrain apply.

Design notes

• Retrieval scoring: BM25 over summary (×3), tags (×2), case id and body, with a CJK bigram tokenizer so Chinese queries work out of the box; results are boosted by verified, use_count and recent last_verified_at, demoted when stale (> 1 year).
• Self-maintenance signals: every query hit increments use_count; log.md feeds memory_distill pattern analysis; lint refreshes nothing silently — every mutation of history goes through human-approved proposals.
• Single-user, local-first: no daemon, no ports; concurrent writes from several agents are serialized by an OS-level byte-range lock (.vault.lock, msvcrt/fcntl — released instantly if the holder crashes), and all file writes are atomic (temp + rename) so readers never see torn files.
• Point-in-time recovery: every vault is its own git repo (created by init, repo-local identity only). Each content write — ingest, apply, lint/distill proposal, suggestion, index rebuild — is auto-committed, so any bad edit can be rolled back with plain git. Query-driven use_count bumps ride along with the next content commit instead of polluting history. Works fully without git; if git is missing, snapshots are silently disabled.

Changelog

• 0.4.6 — Robustness + scriptability round. Hand-written summaries are now normalized to one line and capped at ingest, and both the Index.md table and query output render them one-lined, so a multiline summary can no longer break the index. lint stops flagging retired lessons: ORPHAN/EXPIRED/ LOWCONF skip superseded entries (SECRET/DANGLING still apply), ending recurring proposals that could never be acted on. The redaction scanner exempts lesson-id-shaped words (…-lesson-01, vault file names) so citing pypi-…-lesson-01 no longer trips the PyPI-token pattern. agentbrain snapshot distinguishes committed / clean / failed instead of reporting every git failure as "nothing to commit". lint and doctor gained scripted exit codes (0 clean/healthy, 1 findings/issues, 2 bad scope / uninitialized vault). CLI output reconfigures to UTF-8 when piped, so emoji inside lesson text no longer crash redirects on Chinese-locale Windows. Query results are more token-frugal: three lines per hit (tags and path merged) and top_k clamped to 1–20; a new tag filter (memory_query(tag=…) / agentbrain query --tag …) narrows search to one tag. Index.md dropped its used column, so use-count bumps no longer rewrite the whole index. agentbrain doctor now reports broken lesson files (frontmatter present but unparseable) instead of letting them vanish silently, and lesson-id lookups reject path separators. python -m agentbrain serve works as a PATH-free MCP fallback (the doctor snippet shows both forms), lint findings carry explicit remedies, empty queries and unknown lint scopes are rejected with clear messages, tags past 8 are reported as dropped, oversize lessons (> 4000 chars) get a split hint, and nested locks across two vaults in one thread no longer deadlock. agentbrain.__version__ is now the single version source (pyproject reads it dynamically), and a GitHub Actions matrix (ubuntu/windows × 3.10–3.12) runs the suite. 165 tests (+34).
• 0.4.5 — Signposting + standards round. Agents looking at the owner profile no longer miss the write channel: memory_profile output and the agentbrain://profile resource end with a note that the profile is read-only for agents and changes go through memory_suggest (_suggestions/, owner review), and the memory_suggest tool description states it is the only agent-writable path toward hard rules — closing the gap where an agent concluded "no interface exists" while the tool sat in its context. New rules --agent agentsmd target writes the agents.md open-standard file, so OpenCode, Gemini CLI, Amp and other standard-compliant clients get the discipline block with one command instead of per-client wiring. README MCP examples now use the portable default (~/agentbrain) and note that AGENTBRAIN_VAULT is optional. 131 tests (+4).
• 0.4.4 — Autonomy + upgrade path + robustness round. Agents now ingest lessons autonomously and recall them without being told: MCP tool descriptions (always in the client's context once the server is registered) directly instruct when to query — task start, new subtask/error mid-task, before debugging — and to ingest the moment something is learned, with no user approval needed; the installed rule block and AGENTS.md say the same, and an empty memory_query result now suggests retrying with broader keywords or the other language before concluding nothing is stored. Upgrades no longer strand old versions in existing installs: new agentbrain upgrade refreshes vault templates (AGENTS.md, ONBOARDING.md) — files that match a previously shipped template are updated in place, customized ones are kept and reported; agentbrain rules --write now refreshes an outdated discipline block in place (marker-based, never touching surrounding user content), and --global writes the user-level rule file (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md) so the discipline applies to every project instead of one; doctor reports outdated templates. New agentbrain verify <id> marks a lesson as re-verified today — the remedy for the STALE finding lint reports. memory_ingest warns when a near-duplicate active lesson exists (summary Jaccard ≥ 0.6), so repeated ingests surface immediately instead of waiting for the weekly lint. Redaction false positives fixed: 40-char git SHA-1 digests and Bearer-shaped English prose no longer trip the AWS-secret/Bearer patterns (real secrets still do). Lesson files saved with a UTF-8 BOM by Windows editors (Notepad) are now read correctly instead of silently vanishing from the index; the same BOM tolerance applies to profiles, logs, proposals and MCP resources. Long case ids are truncated (48 chars) instead of crashing with a path-too-long error; verified: 'false' in quoted frontmatter parses as false; lint now also scans tags and case ids for secrets (same scope as ingest); rule-file writes and git snapshot output decoding are hardened. 127 tests.
• 0.4.3 — New agentbrain rules command: registers MCP and the tools become available, but clients only call them if a rule tells them to. One command now installs the memory discipline (query at task start, re-query on new subtasks/errors, ingest at wrap-up with confirmation, secrets never) into the client's own rule file — --agent claude|codex|trae|cursor --write writes CLAUDE.md / AGENTS.md / .trae/rules/ / .cursor/rules/ (Cursor gets an alwaysApply frontmatter block). Existing rule files are never overwritten — the block is appended; re-runs are no-ops via a marker. 95 tests.
• 0.4.2 — Self-service onboarding + hardening: new ONBOARDING.md in every vault routes first-time agents to the right access mode (MCP tools / shell / file-only) — onboarding a new agent is now just "read AGENTS.md"; memory_ingest also scans case_id and tags for credentials (previously only lesson text and summary — a key smuggled into a filename or tag could slip past); hand-edited frontmatter with non-numeric confidence/use_count no longer breaks vault reads (per-field fallback to defaults). 87 tests.
• 0.4.1 — Enforced secret redaction: memory_ingest scans content and summaries for credential-shaped patterns (OpenAI/Anthropic/GitHub/AWS/Slack/Google tokens, Bearer headers, private-key blocks, password=/api_key= assignments) and refuses the write with a placeholder hint — the "secrets never enter the vault" rule is now a mechanism, not just AGENTS.md discipline. ${ENV:VAR} references and teaching examples (sk-xxx) pass through. lint reports SECRET findings for pre-existing lessons (read-only). 84 tests.
• 0.4.0 — Maturity pass: git snapshots (every vault is a self-contained git repo; every content write is an auto-commit you can roll back — repo-local identity, graceful without git), agentbrain doctor one-shot health check (vault, index freshness, lock round-trip, snapshot status, log; prints a copy-paste MCP config with your vault path), agentbrain snapshot manual commit, fool-proof 3-step quickstart, PyPI-ready packaging. 74 tests.
• 0.3.2 — Locking rewrite + edge cases: the vault lock now uses OS-level byte-range locks (msvcrt on Windows, fcntl on POSIX) instead of create-file-and-reclaim — a crashed holder releases instantly (previously all writes failed for up to 60 s) and the stale-reclaim race (two waiters both unlinking and both acquiring) is gone. agentbrain lint --scope tag:x no longer reports false DANGLING for supersede targets outside the scope; case_ids containing glob metacharacters ([, ?, *) no longer collide lesson ids; suggestion files use real YAML frontmatter (titles with colons/newlines used to corrupt it) and atomic writes. 65 tests.
• 0.3.1 — Data-integrity fixes: concurrent same-case ingests no longer overwrite each other (lesson-id allocation moved inside the vault lock); confidence: 0.0 round-trips correctly (was silently coerced to 0.8); lint/distill proposals are written atomically under the lock with collision-free names; merge proposals now keep the more-used lesson as the keeper; duplicate detection pre-tokenizes (O(n²) without re-tokenizing per pair). Session wrap-up rule added to AGENTS.md. 59 tests.
• 0.3.0 — Concurrency & robustness: cross-process/thread vault write lock (.vault.lock, re-entrant, stale-reclaim), atomic writes (temp + rename), apply is now a single transaction; query no longer rebuilds the index once per hit (one rebuild per query); stray non-lesson .md files in Learnings/ are ignored; confidence clamped to [0,1]; unknown mode falls back to index; same-second suggestions no longer overwrite each other. 54 tests.
• 0.2.0 — Owner profile layer (memory_profile / memory_suggest + MCP resources), lint/distill proposals with machine-readable directive blocks, agentbrain apply with cycle/self-supersede/dangling checks.
• 0.1.0 — Initial MVP: vault + frontmatter + CJK-aware BM25 retrieval, MCP server (query/ingest/lint/distill) + CLI, scaffold templates.

Roadmap — maintenance mode

The core promise — a local, token-efficient, agent-shared long-term memory that you own as plain Markdown — is complete and battle-tested in daily use. The project is now in maintenance mode: bug fixes, compatibility with new MCP client versions, and small quality-of-life improvements. Big new subsystems are deliberately out of scope; if a vault ever grows past a few hundred lessons, these are the parked ideas:

• Hybrid fallback search (SQLite FTS5 + local embedding, RRF fusion)
• Keyring-backed ${ENV:...} resolution helper

Done along the way:

• Enforced secret redaction on ingest + SECRET findings in lint
• Git snapshot on every write
• agentbrain apply <proposal> to execute approved consolidations
• Owner profile layer: memory_profile / memory_suggest + MCP resources
• OS-level cross-process vault lock + atomic writes

Development

pip install -e ".[dev]"
pytest

License

Apache-2.0 — see LICENSE.