memdsl 0.9.0

Schema-extensible agent memory as lintable, queryable source code

memdsl

English | 中文

English

Agent memory as source code the agent reads itself.

0.9.0 release line with experimental Edges

Version 0.9.0 adds an opt-in memdsl.workspace.v3 first-class explicit Edge experiment. It does not rewrite the frozen 0.8 contract: legacy relations and authority remain compatible, explicit Edge proposals always require a human, and explicit supersedes is graph-only. Source remains runtime authority; review/audit is a workflow contract, not a non-bypassable grant ledger. See the Phase 6 design and risk matrix and the release-scope evidence freeze.

Stable 0.8 compatibility contract

The 0.8 contract combines the previously unpublished navigation and View work now shipped by 0.9.0; neither 0.7.0 nor 0.8.0 was released separately. Stable public contracts include the v1 compatibility/authority surfaces, Catalog v1, Trace v1, indexed query/search trace, report diagnostics, public rebuildable CompiledWorkspace / compile_workspace, memdsl.workspace.v2, exact use, the generic dialect_mapping capability, ViewContext/ResolvedView, and the explicit opt-in v2 read schemas. Map v1 remains supported throughout the 0.8 line and will not be reconsidered for removal before 1.0.

The software release is 0.9.0, anchored by tag v0.9.0. Package-level release does not promote the opt-in Edge surface to stable, publish the companion paper, or claim that the paper authority runtime is implemented.

The real-world rollout quality of quarantine/strict, dialect-candidate learning, and host-attested principal integration is still experimental and opt-in. Their safety invariants are not experimental: authorization happens before aggregation, incomplete hard-rule evaluation never becomes ALLOW, non-authoritative edges never gain authority, and repair paths remain open. CompiledWorkspace is a public rebuildable handle; its cache/index layout, compiler contract strings, and synthetic timing constants remain implementation details. Workspace-v3 first-class Edges are experimental. Automatic dialect learning, automatic Edge candidates, inferred authority, and cold-history/ incremental compilation are not shipped. Host extraction/sanitization and private schemas/policies/samples remain outside memdsl and are not release gates.

The repository also carries a focused, unpublished position-paper draft, Review-Gated Authority for Persistent Agent Memory. It is a companion research contract, not a description of shipped runtime conformance: memdsl does not yet implement its authority ledger, digest-bound grants, proof objects, live reduction closure, or Verify sinks. Start with the documentation index for the exact software, paper, citation, license, and publication-readiness boundaries.

memdsl started with a retrieval failure. A memory system built on RAG failed to recall a memory it certainly had; a coding agent pointed at the same raw memory files traced it down in a few steps. Same model — the difference was the context contract. The agent knew the memory existed and could keep reading toward it; the retriever got exactly one similarity guess, and a miss looked identical to absence.

memdsl turns that difference into the product. Long-term memory lives in readable, lintable, reviewable .mem source files, and every runtime surface is built for an agent that reads memory itself:

• A bounded memory Catalog (memdsl catalog, MCP memory_catalog / memdsl://catalog, Python build_memory_catalog): paged module/type/subject/ status navigation with item and byte budgets, revision-bound cursors, and no duplicated structured/text representation. It is the recommended session-start surface for large workspaces.
• A compatible memory map (memdsl map, MCP memory_map / memdsl://map): the existing v1 full index remains available for older clients and is not silently changed into Catalog.
• Indexed queries that explain their misses (search_trace in every serialized pack): deterministic lexical postings preserve the v1 EvidencePack ordering and authority lanes while exposing View/source identity, candidate-pool counts, filter-hidden matches, bounded vocabulary suggestions, and safe retry queries. Suggestions never write aliases or let candidate symbols redirect retrieval.
• A bounded relation Trace (memdsl trace, MCP memory_trace, Python trace_memory): deterministic incoming/outgoing/both BFS navigation with relation filters, explicit cycle/back/cross edges, hard depth/node/edge/byte budgets, and revision-bound cursors. Connectivity is navigation, not proof.
• Report-only link diagnostics (memdsl lint, MCP memory_lint and memdsl://status): duplicate ids, ambiguous/wrong-prefix/dangling targets, unknown relations, revision cycles, and supersede forks are explicit. Cycle edges cannot make every participant disappear, forks never select a winner, and default Map/query authority remains v1-compatible.
• Raw source as the floor (memdsl://file/{file_id}): the agent can always drop down and read the actual declarations, the way a coding agent reads code.

The failure mode that started the project now fails loud — the memory exists, a filter hid it, and the trace says exactly that:

$ memdsl query examples/domains/coding -q "force push main" \
    --type coding.bug_pattern --json
  "missing": [
    "no active declarations matched query terms: ['force', 'push', 'main']",
    "1 active declaration(s) matched the query but were excluded by type/subject filters"
  ],
  "search_trace": {
    "excluded_by_filters": [
      {"id": "coding.project_rule:git.no_force_push", ...}
    ], ...
  }

$ memdsl query examples/domains/coding -q "force push main"
MUST
- [coding.project_rule:git.no_force_push] Never force-push the main branch. (exceptions: []) [status=active; runtime_role=constraint; lifecycle={"status":"active"}]

Try it with Claude Code in two minutes

The core library and CLI support Python 3.9+. The MCP extra and memdsl-mcp server require Python 3.10+ because the upstream MCP SDK does.

# Python 3.10+
pip install "memdsl[mcp]"
claude mcp add memdsl -- memdsl-mcp --workspace ~/memory   # or an examples/ dir

The server instructs the agent to read memory_catalog first, continue only with a cursor bound to the same Source/View and filters, treat no_match as a retry signal, use memory_trace only for bounded explicit-relation navigation, and preflight consequential drafts against MUST constraints with memory_check. Legacy memory_map remains registered for compatible clients.

Governed, typed memory

The second pillar: memory you can review like code, without forcing one author's ontology on every user. Version 0.8 preserves the two-layer type architecture introduced in 0.5 and the lifecycle-safe review contract from 0.6, then adds bounded compiled navigation and explicit workspace-v2 Views:

core memory record     claim / evidence / scope / confidence / lifecycle /
                       access policy / relations

domain type system     coding.project_rule / assistant.commitment /
                       writing.voice_preference / your own memory types

The core owns the stable behavioral contract. A workspace owns its vocabulary. A coding agent, personal assistant, and writing system should not have to call the same thing a preference, boundary, or fact just because memdsl's author chose those words.

Two layers

The core layer provides universal fields and five stable runtime roles:

Runtime role	EvidencePack behavior
symbol	Defines a subject that other memories can reference
constraint	Active declarations surface in MUST and participate in compliance
guidance	Active declarations surface in SHOULD
assertion	Active declarations surface in CONTEXT when relevant
question	Active declarations surface in MISSING rather than as fact

Every searchable non-active hit is isolated under PROVISIONAL, regardless of runtime role. Candidate symbols cannot redirect queries, and candidate constraints cannot enter MUST or compliance.

The same authority boundary applies to relations. A supersedes relation can hide a target only when its source is active and its full-id or unique bare target resolves exactly. Candidate, retracted, archived, ambiguous, duplicate, or wrongly prefixed superseders cannot change query, MUST, or compliance. This is the v1 compatibility authority rule retained by 0.8. The opt-in ResolvedView path uses the same fail-safe base and adds only explicit v2 quarantine enforcement.

The core also understands capabilities such as requires_evidence, searchable, temporal, enforceable, guardable, exceptions_recommended, and the explicit review opt-in auto_approvable.

The domain layer defines meaningful memory types and compiles each one to a runtime role. The shipped examples include:

• coding: coding.project_rule, coding.bug_pattern, coding.tool_preference
• personal assistant: assistant.routine, assistant.commitment, assistant.relationship_context
• writing: writing.voice_preference, writing.taboo_topic, writing.style_example

The built-in memdsl.standard@1 pack keeps existing entity, fact, preference, boundary, principle, decision, state, and open_issue workspaces working. Those names are compatibility defaults, not the universal ontology of human memory.

Define a domain vocabulary

A workspace opts into schemas with memdsl.json:

{
  "schema_version": "memdsl.workspace.v1",
  "schemas": ["coding.memschema.json"]
}

The schema defines domain types, fields, roles, and capabilities:

{
  "name": "coding",
  "version": "1",
  "types": {
    "project_rule": {
      "runtime_role": "constraint",
      "required_fields": ["claim", "evidence", "scope"],
      "optional_fields": ["rationale", "owner"],
      "search_fields": ["rationale", "owner"],
      "capabilities": [
        "requires_evidence",
        "searchable",
        "enforceable",
        "guardable",
        "exceptions_recommended"
      ],
      "defaults": {"force": "hard", "status": "active"},
      "allowed_forces": ["hard"],
      "allow_extra_fields": false
    }
  }
}

project_rule becomes the namespaced type coding.project_rule:

module coding.memory

entity Repository.Memdsl {
  canonical_name: "memdsl"
  status: active
}

coding.project_rule git.no_force_push {
  subject: Repository.Memdsl
  claim: "Never force-push the main branch."
  scope: repository("memdsl")
  confidence: high
  exceptions: []
  rationale: "Published history must remain auditable."
  guard {
    when_any: ["push", "git", "main"]
    deny_any: ["force-push", "--force", "--force-with-lease"]
  }
  lifecycle { status: active }
  access_policy {
    readers: [developer, coding_agent]
    writers: [maintainer]
    reviewers: [maintainer]
    export: internal
  }
  evidence {
    source: AGENTS.md
    quote: "Do not force-push unless the user explicitly asks."
  }
}

Strict schemas reject undeclared fields. Unknown memory types fail lint with unknown_memory_type; they do not silently acquire ad-hoc behavior.

Exact imports and workspace-owned dialect

Existing workspaces stay compatible: no manifest and memdsl.workspace.v1 keep legacy global linking. A workspace can explicitly audit or enforce imports with v2:

{
  "schema_version": "memdsl.workspace.v2",
  "schemas": ["workspace-dialect.memschema.json"],
  "linking": {"visibility": "report"}
}

use X is resolved after the whole workspace is indexed, so declaration order does not matter. X is one exact module or one exact active symbol name. There is no wildcard, prefix, alias, or module-over-symbol precedence. Report mode keeps legacy links and emits migration diagnostics; strict mode removes unimported relation, subject-routing, and dialect-routing effects. Strict is never inferred for a v1 workspace.

Dialect remains Source owned by the workspace. A schema type opts in through the generic dialect_mapping capability; the fictional runnable example is examples/dialect/. Only active, public, valid, unambiguous positive mappings route. Candidate, pending, private, ambiguous, and negative mappings do not. A no-match may return a search_trace.dialect_candidate template, but it does not write anything: evidence, proposal, review, approval, and recompilation are still required.

Opt-in quarantine enforcement

Workspace v2 can independently enable the v2 read gate:

{
  "schema_version": "memdsl.workspace.v2",
  "schemas": ["workspace-dialect.memschema.json"],
  "linking": {"visibility": "report"},
  "enforcement": {"mode": "quarantine"}
}

Omitting enforcement, or selecting report, preserves the v1 read behavior. quarantine and strict are explicit opt-ins; they classify Source as authoritative, provisional, quarantined, or excluded and use new v2 query/list/explain/check/Catalog/Trace envelopes. Empty results distinguish no_match, quarantined, unauthorized, compiler_error, and budget limits. Identity-critical errors fail closed, while relation/use/dialect/type errors are quarantined at the smallest safe declaration, file, or revision-family scope. Lint, source edits, proposals, review, and audit remain available for repair.

Map stays a v1 compatibility surface and returns unsupported_view under explicit enforcement. Use Catalog, query, list, explain, or Trace instead. Pending proposals still do not enter a durable View, and candidate/private/ ambiguous dialect mappings never gain routing authority through quarantine.

Install and inspect

The base package supports Python 3.9+:

pip install memdsl                 # or: pip install -e . from a checkout
memdsl --version
memdsl types examples/domains/coding
memdsl lint examples/domains/coding
memdsl catalog examples/domains/coding --json
memdsl map examples/domains/coding
memdsl query examples/domains/coding -q "force push main"
memdsl trace examples/alex decision:aurora.pricing_free_tier --both --depth 2 --json
memdsl check examples/domains/coding \
  -t "push main" \
  -c "git push --force origin main"

memdsl types shows every loaded standard and domain type together with its runtime role, required fields, capabilities, schema version, and source. Use --json when another tool needs to discover the type system.

The custom coding rule behaves exactly like a MUST constraint without any Python code knowing the name coding.project_rule:

MUST
- [coding.project_rule:git.no_force_push] Never force-push the main branch. [status=active; runtime_role=constraint; lifecycle={"status":"active"}]

The compliance check returns BLOCK and cites the same domain declaration. An unguarded or non-enforceable constraint fails safely to NEEDS_REVIEW instead of being guessed at.

Layered query contract

Queries return an EvidencePack, not a flat hit list:

• MUST: applicable constraint declarations
• SHOULD: relevant guidance declarations
• CONTEXT: relevant assertion declarations
• PROVISIONAL: scored non-active searchable declarations
• CONFLICT: declared conflicts among selected memories
• MISSING: relevant question declarations and known gaps

Every item carries a stable declaration id, type, runtime role, capabilities, claim, subject, scope, confidence, lifecycle, access policy, evidence, and source location. Domain semantics stay expressive while runtime behavior stays predictable.

JSON output remains memdsl.evidence_pack.v1; the provisional field added in 0.6 remains additive in 0.8. Scored CONTEXT and PROVISIONAL entries include score and matched_terms, and every declaration item carries explicit lifecycle status and runtime role.

The reference retriever is deliberately lexical and uses a rebuildable inverted term index. Production systems can put BM25, embeddings, graphs, or database indexes behind the same EvidencePack contract.

Navigation: bounded Catalog first, then query and drill down

memdsl exists because agent-driven reading beats blind similarity matching: an agent that knows a memory exists can drill down to it, while a retriever that misses just returns nothing. Four surfaces support that loop:

• Bounded Catalog (memdsl catalog, MCP memory_catalog / memdsl://catalog, Python build_memory_catalog): module summaries filtered by module/type/subject/status. The default page is 20 items / 8192 canonical compact UTF-8 JSON bytes. truncated, next_cursor, returned_items, exact totals, and vocabulary total/truncated metadata are explicit. Cursors bind Source fingerprint, report-only View, filters, order, and representation; changed Source returns cursor_stale.
• Map v1 compatibility (memdsl map, MCP memory_map / memdsl://map, Python build_memory_map): still returns the full serviceable declaration index for existing clients. It is not the recommended large-workspace session-start surface. Under explicit workspace-v2 enforcement it returns unsupported_view; Map and Catalog are navigation, not citation.
• Search trace (search_trace in serialized packs): how the query was interpreted, which indexes ran, the candidate pool before/after filters, and which matching declarations a type/subject filter excluded. A no-match may add bounded lexical vocabulary_suggestions and retry_queries; candidate, ambiguous, or access-restricted vocabulary never becomes an automatic route.
• Bounded Trace (memdsl trace, MCP memory_trace, Python trace_memory): follows explicit resolved relations with deterministic BFS. Defaults are depth 3, 20 nodes, 40 edges, and 8192 bytes. Stateless cursors bind Source/View, anchors, direction, relation filters, depth, and provisional visibility. Back/cycle/cross edges are explicit; graph connectivity is not evidence or proof.

Preflight constraints

Any schema-defined constraint can participate in memdsl check. To run a deterministic guard, its type must declare both enforceable and guardable. Phrase and regex guards produce ALLOW or BLOCK; natural-language-only constraints produce NEEDS_REVIEW. Declared exceptions must be asserted explicitly with --exception.

The standard v0.4 boundary type follows this same generic path. Its old JSON aliases such as boundary_id and applicable_must remain available for client compatibility.

Run the reproducible standard or custom-domain suites:

memdsl eval compliance examples/compliance \
  --cases examples/compliance/cases.jsonl --json

memdsl eval compliance examples/domains/coding \
  --cases examples/domains/coding/cases.jsonl --json

Use as an MCP server

The MCP server requires Python 3.10+:

# Python 3.10+
pip install "memdsl[mcp]"
memdsl-mcp --workspace examples/domains/coding --inspect
memdsl-mcp --workspace ~/memory

The MCP server exposes:

• tools: memory_catalog, memory_map, memory_types, memory_query, memory_trace, memory_check, memory_explain, memory_list, memory_lint, memory_propose, memory_review_list
• resources: memdsl://status, memdsl://catalog, memdsl://map, memdsl://types, memdsl://files, memdsl://file/{file_id}

Agents should read memory_catalog at session start so navigation remains bounded as the workspace grows, then use memory_query and bounded memory_trace for local drill-down. memory_map remains for v1 clients. Before proposing a declaration, call memory_types or read memdsl://types. This discovers the workspace's vocabulary instead of inventing a type or assuming the standard pack is the only valid worldview.

Access can be narrowed with --scopes or MEMDSL_MCP_SCOPES (default: read:summary,read:search,write:candidate). The core represents and transports declaration-level access_policy. For an enforced View, an embedding host may inject a trusted principal and roles only when constructing MemdslMCPService; MCP tool arguments cannot self-assert identity. Filtering happens before counts, vocabulary, diagnostics, graph traversal, and raw-file resources. memdsl does not provide an identity provider, so the host must authenticate the principal and map roles.

Gated writes

Every MCP write is still a proposal. It must use a loaded type, parse, and pass lint against the live workspace. Invalid proposals are rejected; exact pending/approved duplicates return no_op; everything else is either queued or narrowly auto-approved by host-attested policy.

The default remains all-human. Initialize a valid but disabled policy:

memdsl review policy init memory
memdsl review policy show memory
memdsl review policy validate memory

The generated JSON has empty clients/rules and a zero daily limit. To enable automation, a workspace owner must explicitly add auto_approvable to one candidate assertion type, configure an exact-kind rule and trusted host client, set a positive daily limit, and grant the non-default write:auto scope:

memdsl-mcp --workspace memory \
  --scopes read:summary,read:search,write:candidate,write:auto

The built-in workspace_file_quote verifier requires evidence.source to resolve inside a loaded workspace root and evidence.quote to occur exactly in that UTF-8 file. Question, guidance, constraint, symbol, active, global, warned, destructive, or unverified proposals always stay queued. Without write:auto, a valid policy runs in shadow posture: the write remains pending while eligible_route records what policy would have done.

Human operations and quality feedback remain explicit:

memdsl review list memory
memdsl review approve memory PROPOSAL_ID --into memory/approved.mem
memdsl review reject memory PROPOSAL_ID --reason "not durable"
memdsl review digest memory
memdsl review stats memory
memdsl review audit memory PROPOSAL_ID --verdict confirm

In-process hosts use authoritative paths:

from memdsl import (
    ProposalContext,
    ReviewStore,
    Workspace,
    load_policy,
    staging_dir_for,
)

workspace = Workspace.load(["memory"])
store = ReviewStore(staging_dir_for(["memory"]))
policy = load_policy(store.staging_dir, registry=workspace.registry)
assert policy is not None
context = ProposalContext(client_id="mcp-client")

result = store.submit(
    ["memory"],
    proposal_source,
    policy=policy,
    context=context,
    write_auto_granted=True,
)

Automatic approval reloads and fingerprints memory, manifest, and schema inputs, re-verifies evidence, enforces a finite UTC daily limit, writes only to a non-symlink .mem target inside the primary workspace root, and never uses force. Route assessments, decisions, post-review results, digest cursors, and no-op events remain append-only.

A post-review flag does not silently delete memory. Promotion, revision, and retraction require a new human-reviewed declaration with a new id and supersedes (optionally revision_of) pointing to the old declaration. The successor must be lifecycle active before the relation has authority; the old declaration does not need an in-place status: superseded rewrite. ReviewStore does not create Git commits; hosts may add Git integration without making core correctness depend on it.

What's in the box

• A domain-neutral .mem record with provenance, scope, confidence, lifecycle, access policy, relations, and stable ids.
• An extensible, namespaced .memschema.json type system and workspace manifest.
• A backward-compatible standard type pack for pre-v0.5 workspaces.
• A schema-driven linter, layered query executor, explainer, and deterministic Compliance Gate.
• Type discovery through CLI and MCP.
• A default-safe, host-attested review pipeline with human queueing, deterministic narrow auto-approval, no-op detection, digest/stats, and append-only audit.
• Reproducible compliance benchmarks and coding/assistant/writing domain packs.
• Fictional standard examples for Alex and Mira plus a deliberately broken linter workspace.

Full grammar and semantics are in docs/SPEC.md. In-process hosts should also read docs/PUBLIC_API.md and docs/UPGRADING.md. The review-policy security contract is in docs/DESIGN_review_policy.md. The documentation index keeps the practical source/compiled-view design, focused paper, claim ledger, reproducibility record, readiness audit, citation, and license boundaries together.

What memdsl is not

• It is not a replacement for retrieval/extraction systems such as Mem0, Zep, Graphiti, or LangMem. It is the governed source format and behavioral contract above those systems.
• It is not a universal taxonomy of people or memory. Domain owners define their own types.
• It is not a semantic policy oracle. Constraints that cannot be evaluated deterministically remain NEEDS_REVIEW.
• It is not an unrestricted automatic memory writer. Only explicitly opted-in candidate assertions with trusted identity and verified evidence can be auto-approved; every higher-risk or uncertain proposal remains human reviewed.

Current evidence and limits

Early evidence from the private single-user system this project was extracted from showed higher retrieval precision than that system's tuned RAG baseline on its internal evaluation. That result is encouraging, but it is not evidence that one ontology generalizes to everyone. The v0.8 architecture makes that limitation explicit: memdsl standardizes the record and runtime contract while letting each domain own its vocabulary.

The bundled suites are contract-level, deterministic tests. Cross-model behavioral claims still require separately recorded model runs.

Roadmap

• Two-layer core + extensible domain type system — shipped in v0.5
• Lifecycle-safe provisional serving + tiered review policy — shipped in v0.6
• Public rebuildable CompiledWorkspace + bounded Catalog pagination/budgets — stable/public 0.8 contract shipped by 0.9.0
• Indexed query + bounded Trace + exact use/report/strict + reviewed Dialect — stable/public 0.8 contract shipped by 0.9.0
• Opt-in quarantine enforcement with explicit v2 envelopes — stable/public schema/safety contract; real rollout quality remains experimental
• Opt-in first-class reviewed Edge experiment — shipped as experimental in 0.9.0 after an anonymous exploratory Pilot; relation/evidence narrowing remains ADJUST, not a formal empirical validation
• Cold history/incremental compilation — deferred until representative scale, SLO, and production bottleneck evidence exists
• Schema package distribution, dependency/version constraints, and migrations
• Richer field validators and domain-defined diagnostic rules
• Pluggable retrieval backends behind the EvidencePack contract
• Identity-provider adapters for enforcing represented access policies
• Semantic conflict reviewers and per-client/per-rule quotas, gated by measured shadow/sample/post-review evidence

License

Code: MIT. Specification (docs/SPEC.md): CC-BY-4.0. The focused paper artifacts are separately licensed under CC BY 4.0.

---

中文

0.9.0 正式软件版本中的实验 Edge

0.9.0 新增显式 opt-in 的 memdsl.workspace.v3 一等 Edge 实验，不改写已经冻结的 0.8 合同：legacy relation 与 authority 保持兼容，Edge proposal 永远进入人工 审核，显式 supersedes 在本阶段仅进入 graph。当前 runtime authority 仍来自 Source；review/audit 只是 workflow contract，不是不可绕过的授权账本。详见 Phase 6 设计与风险矩阵 与 发布范围证据冻结。

把 Agent 记忆写成 agent 自己会去读的源代码。

Stable 0.8 兼容合同

0.8 合同把此前未单独发布的导航能力与 View 能力统一后由 0.9.0 承载；0.7.0 与 0.8.0 均未单独发布。正式公共合同包括 Source/review authority、公共且可重建的 CompiledWorkspace / compile_workspace、v1 兼容 surface、Catalog v1、Trace v1、indexed query/search trace、report diagnostics、memdsl.workspace.v2、 exact use、通用 dialect_mapping capability、ViewContext/ResolvedView 与 显式 opt-in 的 v2 read schemas。Map v1 在整个 0.8 line 继续保留，不早于 1.0 才重新评估删除。

正式软件版本为 0.9.0，由 tag v0.9.0 锚定。软件版本发布不会把 opt-in Edge 提升为 stable，也不代表配套论文已经发表，更不代表论文 authority runtime 已实现。

quarantine/strict 的真实 rollout 质量、dialect candidate 学习闭环和宿主证明 principal 的集成仍是 experimental/opt-in；但 authorization-before-aggregation、 hard rule 不完整时不得 ALLOW、非权威 edge 不得获得 authority、repair lane 必须 可用等安全不变量不是实验性的。CompiledWorkspace 是公共、可重建 handle；其 cache/index 布局、compiler contract 字符串和 synthetic timing 仍是内部事实。 workspace-v3 一等 Edge 保持 experimental；自动方言学习、自动 Edge 候选、inferred authority 与冷历史/增量编译尚未发布。宿主 extractor/sanitizer、私人 schema/policy/ 样本不属于 memdsl，也不构成 release gate。

仓库还并列保存一份未发表的聚焦立场论文草稿 Review-Gated Authority for Persistent Agent Memory。 它是配套研究合同，不代表当前 runtime 已符合论文：memdsl 尚未实现论文中的 authority ledger、digest-bound grants、proof objects、live reduction closure 或 Verify sinks。软件、论文、引用、许可和发表就绪边界以 文档索引 为准。

memdsl 起源于一次召回失败。一个基于 RAG 的记忆系统没能召回一条确实存在的 记忆；而一个 coding agent 拿到同样的原始记忆文件，几步就把它找了出来。 同一个模型——差别在上下文契约：agent 知道那条记忆存在，可以一层层读过去； 检索器只有一次相似度猜测，而且"没命中"和"不存在"看起来一模一样。

memdsl 把这个差别做成了产品。长期记忆保存在可读、可 lint、可审查的 .mem 源文件里，每个运行时 surface 都为"agent 自己读记忆"而设计：

• 有界记忆目录 Catalog（memdsl catalog、MCP memory_catalog / memdsl://catalog、Python build_memory_catalog）：按 module/type/subject/ status 分页导航，具备 item/byte 硬预算、绑定 revision 的稳定 cursor，并避免 structured/text 双份返回；这是大型 workspace 推荐的会话起始 surface。
• 兼容记忆地图（memdsl map、MCP memory_map / memdsl://map）：旧的 v1 全量索引继续保留，不会被原地改成 Catalog。
• 带倒排索引、会解释 miss 的查询（序列化 pack 中的 search_trace）： 确定性的词法 postings 保持 v1 EvidencePack 排序与 authority lane，同时返回 View/source、候选池、filter-hidden match、受限词汇建议和安全 retry query；建议 不会写 alias，也不会让 candidate symbol 改写路由。
• 有界关系 Trace（memdsl trace、MCP memory_trace、Python trace_memory）：确定性的 incoming/outgoing/both BFS、relation filter、显式 cycle/back/cross edge、depth/node/edge/byte 硬预算与 revision-bound cursor。 连通性只是导航，不是证明。
• 原始源码兜底（memdsl://file/{file_id}）：agent 随时可以下钻去读 真正的声明，就像 coding agent 读代码一样。

当年启动这个项目的那种失败，现在会大声报错——记忆存在，是过滤器藏了它， trace 会把话说明白：

$ memdsl query examples/domains/coding -q "force push main" \
    --type coding.bug_pattern --json
  "missing": [
    "no active declarations matched query terms: ['force', 'push', 'main']",
    "1 active declaration(s) matched the query but were excluded by type/subject filters"
  ],
  "search_trace": {
    "excluded_by_filters": [
      {"id": "coding.project_rule:git.no_force_push", ...}
    ], ...
  }

$ memdsl query examples/domains/coding -q "force push main"
MUST
- [coding.project_rule:git.no_force_push] Never force-push the main branch. (exceptions: []) [status=active; runtime_role=constraint; lifecycle={"status":"active"}]

两分钟在 Claude Code 里跑起来

底层库和 CLI 支持 Python 3.9+。由于上游 MCP SDK 的要求，MCP extra 和 memdsl-mcp server 需要 Python 3.10+。

# Python 3.10+
pip install "memdsl[mcp]"
claude mcp add memdsl -- memdsl-mcp --workspace ~/memory   # 或任意 examples/ 目录

server 会指示 agent 先读 memory_catalog，只在 Source/View 和过滤条件不变时 继续 cursor，把 no_match 当作重试信号，只用 memory_trace 做有界的显式关系 导航，并在产出重要草稿前用 memory_check 对照 MUST 约束做预检。旧客户端仍可调用 memory_map。

受治理的类型化记忆

第二根支柱：像审代码一样审记忆，但不要求所有用户接受作者的一套世界观。 v0.8 保留 v0.5 引入的两层类型架构与 v0.6 的 lifecycle 安全审核合同，并加入 有界编译导航与显式 workspace-v2 View：

底层通用记录        claim / evidence / scope / confidence / lifecycle /
                    access policy / relations

上层领域类型系统    coding.project_rule / assistant.commitment /
                    writing.voice_preference / 用户自定义类型

底层负责稳定的行为契约，workspace 负责自己的词汇。编程 Agent、个人助理和 写作系统不应该仅仅因为 memdsl 作者选择了 preference、boundary、fact 这些词，就被迫用同一套分类描述自己的记忆。

两层架构

底层提供通用字段和五种稳定的 runtime role：

Runtime role	在 EvidencePack 中的行为
symbol	定义其他记忆可以引用的主体
constraint	只有 active 声明进入 MUST 并参与合规检查
guidance	只有 active 声明进入 SHOULD
assertion	只有 active 声明相关时进入 CONTEXT
question	只有 active 声明进入 MISSING，而不是伪装成事实

任何 searchable 的非 active 命中都只进入 PROVISIONAL，不论 runtime role。 candidate symbol 不能重定向查询，candidate constraint 不能进入 MUST 或 compliance。

同一 authority 边界也适用于关系。只有 source 为 active，且 full id 精确匹配或 bare ref 唯一解析时，supersedes 才能隐藏 target。candidate、retracted、 archived、歧义、重复或错误 kind 前缀的 superseder 都不能改变 query、MUST 或 compliance。这是 0.8 继续保留的 v1 correctness/security authority 规则。 ResolvedView 复用这个安全基础，只在显式 v2 opt-in 时增加 quarantine enforcement。

底层还识别 requires_evidence、searchable、temporal、 enforceable、guardable、exceptions_recommended，以及显式审核 opt-in 的 auto_approvable 等 capability。

上层定义领域真正关心的 memory type，并把每种类型编译到一个稳定 role。 仓库自带三个示例领域：

• 编程：coding.project_rule、coding.bug_pattern、 coding.tool_preference
• 个人助理：assistant.routine、assistant.commitment、 assistant.relationship_context
• 写作：writing.voice_preference、writing.taboo_topic、 writing.style_example

内置的 memdsl.standard@1 兼容包会继续加载 entity、fact、 preference、boundary、principle、decision、state、 open_issue 等旧类型。它们是向后兼容的默认词汇，不是人类记忆的唯一分类法。

定义自己的领域词汇

workspace 通过 memdsl.json 引入 schema：

{
  "schema_version": "memdsl.workspace.v1",
  "schemas": ["coding.memschema.json"]
}

.memschema.json 定义类型的字段、role 和 capability：

{
  "name": "coding",
  "version": "1",
  "types": {
    "project_rule": {
      "runtime_role": "constraint",
      "required_fields": ["claim", "evidence", "scope"],
      "optional_fields": ["rationale", "owner"],
      "search_fields": ["rationale", "owner"],
      "capabilities": [
        "requires_evidence",
        "searchable",
        "enforceable",
        "guardable",
        "exceptions_recommended"
      ],
      "defaults": {"force": "hard", "status": "active"},
      "allowed_forces": ["hard"],
      "allow_extra_fields": false
    }
  }
}

加载后，project_rule 的完整名字是 coding.project_rule：

coding.project_rule git.no_force_push {
  subject: Repository.Memdsl
  claim: "绝不对 main 分支执行 force-push。"
  scope: repository("memdsl")
  confidence: high
  exceptions: []
  rationale: "已发布的历史必须保持可审计。"
  guard {
    when_any: ["push", "git", "main"]
    deny_any: ["force-push", "--force", "--force-with-lease"]
  }
  lifecycle { status: active }
  access_policy {
    readers: [developer, coding_agent]
    writers: [maintainer]
    reviewers: [maintainer]
    export: internal
  }
  evidence {
    source: AGENTS.md
    quote: "除非用户明确要求，否则不要 force-push。"
  }
}

严格 schema 会拒绝未声明字段。未知类型会产生 unknown_memory_type lint 错误，不会悄悄获得一套临时行为。

精确 use 与 workspace 自有方言

旧 workspace 保持兼容：没有 manifest 或使用 memdsl.workspace.v1 时仍是 legacy 全局链接。只有显式使用 v2，才会进入 report 或 strict：

{
  "schema_version": "memdsl.workspace.v2",
  "schemas": ["workspace-dialect.memschema.json"],
  "linking": {"visibility": "report"}
}

compiler 会先建立整库 module/symbol 表，再解析 use X，所以声明顺序不影响结果。 X 只能精确命中一个 module 或一个 active symbol name；不支持通配符、前缀、 alias，也不设置 module 优先或 symbol 优先。report 保留 legacy 链接并给迁移诊断； strict 会去掉未导入 relation、subject 路由和 dialect 路由的效果。v1 绝不会因为 出现了 use 就被暗中切成 strict。

方言仍是 workspace 自己拥有、可版本化、可审核的 Source。schema type 通过通用 dialect_mapping capability opt in，仓库提供了完全虚构、可运行的 examples/dialect/。只有 active、公开、结构有效、无歧义的 positive mapping 才会路由；candidate、pending、private、ambiguous 和 negative mapping 都不会。 no-match 可以返回 search_trace.dialect_candidate 模板，但不会自动写 Source； 仍需 evidence、proposal、人工审核、批准和重新编译。

显式 opt-in 的 quarantine enforcement

workspace v2 可以独立开启 v2 读取 gate：

{
  "schema_version": "memdsl.workspace.v2",
  "schemas": ["workspace-dialect.memschema.json"],
  "linking": {"visibility": "report"},
  "enforcement": {"mode": "quarantine"}
}

省略 enforcement 或选择 report，仍保持 v1 读取行为。quarantine 与 strict 必须显式 opt in；它们把 Source 分成 authoritative、provisional、 quarantined、excluded，并使用新的 query/list/explain/check/Catalog/Trace v2 envelope。空结果会明确区分 no_match、quarantined、unauthorized、 compiler_error 与预算不足。identity-critical 错误 fail closed，relation/use/ dialect/type 错误则尽量只隔离最小安全的 declaration、文件或 revision family。 lint、直接编辑 Source、proposal、review 和 audit 修复通道仍然开放。

Map 保持 v1 兼容面，在显式 enforcement 下返回 unsupported_view；应改用 Catalog、query、list、explain 或 Trace。pending proposal 仍不会进入 durable View，candidate/private/ambiguous dialect 也不会借 quarantine 获得路由权。

安装与类型发现

基础包支持 Python 3.9+：

pip install memdsl                 # 或在 checkout 中运行 pip install -e .
memdsl --version
memdsl types examples/domains/coding
memdsl lint examples/domains/coding
memdsl catalog examples/domains/coding --json
memdsl map examples/domains/coding
memdsl query examples/domains/coding -q "force push main"
memdsl trace examples/alex decision:aurora.pricing_free_tier --both --depth 2 --json
memdsl check examples/domains/coding \
  -t "push main" \
  -c "git push --force origin main"

memdsl types 会列出已加载的标准类型和领域类型，以及各自的 runtime role、 必填字段、capability、schema 版本和来源。需要给其他工具消费时可加 --json。

上面的自定义类型不需要在 Python 代码里写死 coding.project_rule，依然会作为 MUST 约束出现：

MUST
- [coding.project_rule:git.no_force_push] Never force-push the main branch. [status=active; runtime_role=constraint; lifecycle={"status":"active"}]

合规检查会返回 BLOCK 并引用同一个领域声明。没有 guard 或不具备执行能力的 约束会安全返回 NEEDS_REVIEW，不会假装已经理解。

分层查询契约

查询返回的是 EvidencePack，而不是扁平命中列表：

• MUST：适用的 constraint
• SHOULD：相关的 guidance
• CONTEXT：相关的 assertion
• PROVISIONAL：带分数的非 active searchable 声明
• CONFLICT：已声明的冲突
• MISSING：相关的 question 和已知信息缺口

每一项都携带稳定 id、类型、runtime role、capability、claim、subject、scope、 confidence、lifecycle、access policy、evidence 和源码位置。这样领域语义可以扩展， 运行时行为仍然稳定。

JSON 输出仍为 memdsl.evidence_pack.v1；v0.6 加入的 provisional 在 v0.8 继续保持 additive。带分数的 CONTEXT 和 PROVISIONAL 条目都包含 score、 matched_terms，每个 declaration item 还明确携带 lifecycle status 和 runtime role。

参考实现使用可重建的词法倒排索引。生产系统可以在同一个 EvidencePack 契约 后面替换成 BM25、embedding、图或数据库索引。

导航：先读有界 Catalog，再查询和下钻

memdsl 的出发点是"让 agent 自己读记忆"胜过"让相似度匹配替 agent 决定"： 知道某条记忆存在的 agent 可以一层层钻下去，而检索器一旦没命中就什么都没有。 四个 surface 支撑这个闭环：

• 有界 Catalog（memdsl catalog、MCP memory_catalog / memdsl://catalog、Python build_memory_catalog）：按 module 汇总，并可按 module/type/subject/status 过滤。默认 20 items / 8192 canonical compact UTF-8 JSON bytes；truncated、next_cursor、returned_items、精确 total 和词表 total/truncated metadata 都显式返回。cursor 绑定 Source fingerprint、 report-only View、过滤条件、顺序和 representation；Source 改变会返回 cursor_stale。
• Map v1 兼容面（memdsl map、MCP memory_map / memdsl://map、Python build_memory_map）：继续为旧客户端返回全量可服务声明索引，但不再推荐作为 大型 workspace 的会话起始 surface。显式 workspace-v2 enforcement 下会返回 unsupported_view。Map 与 Catalog 都只是导航，不是引用来源。
• 检索痕迹（序列化 pack 中的 search_trace）：记录查询被如何解释，以及 使用了哪些 index、过滤前后候选池，以及 type/subject filter 排除了哪些本来 匹配的声明。no-match 可返回有界 vocabulary_suggestions 与 retry_queries；candidate、歧义或 access-restricted 词汇不会成为自动路由。
• 有界 Trace（memdsl trace、MCP memory_trace、Python trace_memory）：沿已解析的显式关系做确定性 BFS。默认 depth 3、20 nodes、 40 edges、8192 bytes；cursor 绑定 Source/View、anchor、direction、relation、 depth 与 provisional visibility。back/cycle/cross edge 显式返回，但连通性不是 evidence 或 proof。
• 只报告、不隔离的链接诊断（memdsl lint、MCP memory_lint 与 memdsl://status）：duplicate id、歧义/错误前缀/dangling target、未知 relation、 revision cycle 和 supersede fork 都会显式出现。cycle edge 不会让参与节点全部 消失，fork 不会暗选 winner，而默认 Map/query authority 仍保持 v1 兼容。

执行前约束检查

任何被 schema 编译为 constraint 的类型都可以参与 memdsl check。要执行 确定性的 guard，该类型必须同时声明 enforceable 和 guardable。短语与正则 guard 返回 ALLOW 或 BLOCK；只有自然语言的约束返回 NEEDS_REVIEW。 例外必须由调用方通过 --exception 显式声明。

v0.4 的标准 boundary 也走同一条通用链路。boundary_id、 applicable_must 等旧 JSON 字段仍作为兼容别名保留。

memdsl eval compliance examples/compliance \
  --cases examples/compliance/cases.jsonl --json

memdsl eval compliance examples/domains/coding \
  --cases examples/domains/coding/cases.jsonl --json

作为 MCP server 使用

MCP server 需要 Python 3.10+：

# Python 3.10+
pip install "memdsl[mcp]"
memdsl-mcp --workspace examples/domains/coding --inspect
memdsl-mcp --workspace ~/memory

MCP server 提供：

• tools：memory_catalog、memory_map、memory_types、memory_query、memory_trace、memory_check、 memory_explain、memory_list、memory_lint、memory_propose、 memory_review_list
• resources：memdsl://status、memdsl://catalog、memdsl://map、memdsl://types、 memdsl://files、memdsl://file/{file_id}

Agent 应该在会话开始时读取 memory_catalog，让导航成本不随 workspace 总量 线性膨胀，再用 memory_query 与有界 memory_trace 局部下钻；memory_map 继续服务 v1 客户端。在提出新声明前，应该先调用 memory_types 或读取 memdsl://types，发现当前 workspace 的词汇，而不是自己 发明类型，也不是默认标准兼容包就是唯一世界观。

可通过 --scopes 或 MEMDSL_MCP_SCOPES 缩小 MCP 权限范围，默认值为 read:summary,read:search,write:candidate。底层会表示、验证并传输声明级 access_policy。在 enforced View 中，嵌入式宿主只能在构造 MemdslMCPService 时注入可信 principal 与 roles；MCP tool 参数不能自报身份。 过滤发生在 count、vocabulary、diagnostics、graph traversal 和 raw-file resource 之前。memdsl 不提供身份提供商；宿主仍需认证 principal 并映射 roles。

受审查的写入

每次 MCP 写入仍然先成为 proposal。新声明必须使用已加载类型，能够解析，并 通过当前 workspace 的 lint。非法 proposal 会被拒绝；与 pending/approved 内容 完全相同的提交返回 no_op；其余 proposal 要么排队，要么由宿主证明过的窄 策略自动批准。

默认仍然全部人审。先生成一份合法但禁用自动批准的 policy：

memdsl review policy init memory
memdsl review policy show memory
memdsl review policy validate memory

生成的 JSON 里 trusted clients 和 rules 都为空，日限额为 0。要开启自动化， workspace owner 必须给一个 candidate assertion 类型显式加入 auto_approvable，配置精确 kind 规则和可信宿主 client，设置正数日限额， 再授予默认不包含的 write:auto scope：

memdsl-mcp --workspace memory \
  --scopes read:summary,read:search,write:candidate,write:auto

内置 workspace_file_quote verifier 要求 evidence.source 解析到已加载 workspace root 内，且 evidence.quote 必须逐字出现在该 UTF-8 文件中。 question、guidance、constraint、symbol、active、global、带 warning、带破坏性 关系或证据未验证的 proposal 始终排队。没有 write:auto 时，合法 policy 处于 shadow 姿态：写入仍 pending，但 eligible_route 会记录策略原本会怎么做。

人工操作与质量反馈保持显式：

memdsl review list memory
memdsl review approve memory PROPOSAL_ID --into memory/approved.mem
memdsl review reject memory PROPOSAL_ID --reason "not durable"
memdsl review digest memory
memdsl review stats memory
memdsl review audit memory PROPOSAL_ID --verdict confirm

进程内宿主应传入权威 workspace paths：

from memdsl import (
    ProposalContext,
    ReviewStore,
    Workspace,
    load_policy,
    staging_dir_for,
)

workspace = Workspace.load(["memory"])
store = ReviewStore(staging_dir_for(["memory"]))
policy = load_policy(store.staging_dir, registry=workspace.registry)
assert policy is not None
context = ProposalContext(client_id="mcp-client")

result = store.submit(
    ["memory"],
    proposal_source,
    policy=policy,
    context=context,
    write_auto_granted=True,
)

自动批准前会重新加载并 fingerprint 记忆、manifest 和 schema，重新验证证据， 执行有限 UTC 日配额，只写入主 workspace root 内的非 symlink .mem 目标， 且绝不使用 force。路由 assessment、decision、post-review 结果、digest cursor 和 no-op 事件都保持 append-only。

post-review flag 不会静默删除记忆。晋升、修订和撤销都要提交一条新 id 的 人工审核声明，用 supersedes（可同时用 revision_of）指向旧声明。successor 必须是 lifecycle active 才获得关系 authority；旧声明不需要原地改成 status: superseded。 ReviewStore 不创建 Git commit；宿主可以增加 Git 集成，但核心正确性不依赖 Git。

包里有什么

• 与领域无关的 .mem 通用记录：来源、scope、confidence、lifecycle、 access policy、relations 和稳定 id。
• 可扩展、带 namespace 的 .memschema.json 类型系统和 workspace manifest。
• 面向 v0.5 之前 workspace 的标准兼容类型包。
• schema 驱动的 linter、分层查询、explain 和确定性 Compliance Gate。
• CLI 与 MCP 的类型发现能力。
• 默认安全、宿主证明的审核链路：人工队列、确定性的窄范围自动批准、no-op 检测、digest/stats 和 append-only audit。
• 可复现合规 benchmark，以及 coding、assistant、writing 三个领域包。
• Alex、Mira 两个虚构标准示例和一个故意损坏的 lint 示例。

完整语法与语义见 docs/SPEC.md。进程内宿主还应阅读 docs/PUBLIC_API.md 和 docs/UPGRADING.md。 审核策略的安全合同见 docs/DESIGN_review_policy.md。 文档索引 把实践设计、聚焦论文、claim ledger、 复现记录、就绪审计、引用与许可边界集中在一起。

memdsl 不是什么

• 它不是 Mem0、Zep、Graphiti、LangMem 等检索/抽取系统的替代品，而是位于 这些系统之上的受治理源格式和行为契约。
• 它不是人或记忆的通用分类法；领域所有者定义自己的类型。
• 它不是语义策略裁判；无法确定性执行的约束会保持 NEEDS_REVIEW。
• 它不是不受限制的自动记忆写入器。只有显式 opt-in、身份可信且证据已验证的 candidate assertion 才可能自动批准；所有高风险或不确定 proposal 仍由人审。

当前证据与边界

这个项目来自一个私有的单用户系统。该系统的内部评测曾显示，DSL 结构化检索的 precision 高于它自己的调优 RAG baseline。这个结果值得继续验证，但不能证明一套 ontology 对所有人都通用。v0.8 正面承认这个限制：memdsl 标准化记录结构和运行 契约，把领域词汇的所有权交还给使用者。

仓库内置的是确定性的契约级测试。跨模型行为结论仍需单独记录真实模型运行。

Roadmap

• 两层 core + 可扩展领域类型系统 — v0.5 已完成
• lifecycle 安全的 provisional 服务 + 分级审核策略 — v0.6 已完成
• 公共可重建 CompiledWorkspace + 有界 Catalog 分页/预算 — 0.9.0 已发布的 stable/public 0.8 合同
• 倒排查询 + 有界 Trace + exact use/report/strict + 受审核 Dialect — 0.9.0 已发布的 stable/public 0.8 合同
• 显式 v2 envelope 下的 opt-in quarantine enforcement — schema/安全合同 stable/public，真实 rollout 质量仍为 experimental
• 显式 opt-in 的一等可审核 Edge 实验 — 0.9.0 以 experimental 状态承载该能力，依据匿名探索性 Pilot 开启；relation/evidence 收窄结论仍为 ADJUST，不是正式实证验证
• 冷历史/增量编译 — 等待代表性规模、SLO 与生产瓶颈证据
• Schema 包分发、依赖/版本约束和迁移机制
• 更丰富的字段验证器和领域诊断规则
• EvidencePack 后面的可插拔检索后端
• 执行 access policy 的身份提供商适配器
• 基于 shadow/sample/post-review 数据的语义冲突 reviewer 与 per-client/per-rule 配额

许可证

代码：MIT。规范（docs/SPEC.md）：CC-BY-4.0。 聚焦论文材料另按 CC BY 4.0 授权。