MyMCP

MyMCP is the repository and Python host package for an experimental local MCP server. Released 0.2.0 identifies the host, application, package, and official tracked client as MyMCP/mymcp. It hosts Mnemosyne, the user-governed AI memory domain, through a bundled plugin.

Its intended direction is a local, client-neutral MCP host and governance gateway that composes narrowly scoped integrations behind one machine-local endpoint. Mnemosyne supplies the current notebook-like memory capability while retaining its explicit user-governed boundaries.

Current Status

This repository currently contains a FastAPI-based MCP host with the Mnemosyne memory domain built in. Phase 1 provides kind-qualified contracts, ActivatedTool/PluginContribution composition, trusted internal effect and consent metadata, host-owned public bindings and list_tools, an immutable HostRuntime with an opaque generation, explicit build_production_runtime, and runtime-bound MCP dispatch through create_app(runtime).

Phase 2's declaration/parity boundary and vertical Mnemosyne extraction are implemented. The bundled plugin owns its inert manifest, trusted adapter, configuration, memory domain, and memory_* MCP adapters under mymcp/plugins/mnemosyne/. Manifest-v1 parsing remains strict (JSON/UTF-8, duplicate-key, and 64 KiB bound), and bootstrap validates exact manifest/definition/contribution parity before generation construction.

The plugin's trusted in-process Mnemosyne 0.1.0 adapter supplies canonical registrations. Host bootstrap reads only the fixed packaged resource, validates it against the plugin definition and selected gate-controlled contribution, then constructs a runtime. create_production_app() is the local convenience factory and supported Uvicorn factory target. Ordinary imports are side-effect free; no global app, startup/methods modules, static ToolIntegration, or dynamic discovery are used.

TRACK_034 delivered the MyMCP 0.2.0 public-host cutover: server/application identity, package metadata, the canonical repository, and the tracked OpenCode policy use MyMCP/mymcp. The prior repository URL redirects to the canonical repository. Routes and protocol remain unchanged. Mnemosyne permanently retains plugin mnemosyne, memory_* Tools, MNEMOSYNE_* configuration, ~/.mnemosyne, and its storage, record, logging, and consent semantics. Repository/origin/history/ tag/placeholder checks, the tracked and ignored OpenCode migration with Claude exclusion, the normal endpoint and final client reconnect, and isolated approved-once and rejected/no-Tools-call checks are complete. The public, non-draft, non-prerelease GitHub release MyMCP 0.2.0: public-host cutover is tagged mymcp-v0.2.0 at c2852bc and contains one wheel, mymcp-0.2.0-py3-none-any.whl, with matching GitHub/local SHA-256 531cc9a603d16399b12650fd09d3bc76f43b4d5d1b15fed0377a6197c820e3e7. External installation/activation/isolation, lifecycle, and gateway governance remain deferred. See docs/PLUGIN_ARCHITECTURE.md for the target and migration boundaries.

Implemented tools:

• list_tools — reports the server version and lists the exposed tools
• memory_recall — retrieves bounded, relevant, user-approved JSON memory records from one allowlisted local scope directory
• memory_list — completely and deterministically inventories valid memories in one bounded scope or canonical container without returning record content
• memory_inspect — returns one exact canonical or legacy memory selected by a versioned structured reference
• memory_remember — when explicitly enabled, validates and atomically persists one approved canonical version-2 memory
• memory_revise — when independently enabled, revision-checks and atomically replaces the complete caller-mutable state of one exact canonical version-2 memory
• memory_archive — when explicitly enabled, revision-checks and archives one exact canonical version-2 memory
• memory_restore — when explicitly enabled, revision-checks and restores one exact archived canonical version-2 memory
• memory_forget — when independently enabled, permanently deletes one exact archived canonical version-2 memory at its expected revision

This is not yet a full memory substitute. A memory_recall request contains a free-form query, exactly one high-level scope (self, relationship, preference, practice, project, or knowledge), and optionally 1–10 unique free-form tags. Scope selects one fixed local directory. Query terms and tags rank valid records from that directory; recall never searches another scope or accepts a client-supplied path. The Tool schema publishes scope as an explicit string enum so clients can discover the complete vocabulary.

Matching calls return a normal Tool result with status: ok and at most five memory records. Results include an inspect-compatible versioned reference, the record ID, scope, title, content, tags, and matched terms/tags, but never include filesystem paths or internal scores. A legacy reference contains schema version, scope, and ID. A canonical reference additionally contains namespace ID and the nullable collection ID. An absent scope directory or no relevant record returns {"status":"no_matches","memories":[]}. Invalid arguments retain stable Tool errors with code invalid_query, invalid_scope, or invalid_tags; unreadable or excessive sources return memory_source_unavailable or candidate_limit_exceeded Tool errors.

Recall remains read-only. It does not create, update, delete, or automatically extract memory, and it does not persist recall requests. Calls remain visible through the MCP client's existing Tool-call/session representation.

Listing Memory

memory_list is registered by default as a read-only Tool for complete bounded discovery. It is distinct from both other read operations:

• recall ranks records by a query, excludes non-matches and archived canonical records, and returns at most five results;
• listing enumerates every valid selected record without a query or relevance score, includes archived canonical records, and returns compact metadata in deterministic pages;
• inspection retrieves the complete user-visible contents of one exact listed or recalled reference.

Every list request requires exactly one scope. Its Tool schema exposes scope, namespace_id, collection_id, page_size, and cursor as top-level object properties for clients that do not project properties nested in composition branches. Scope uses the same explicit six-value string enum as recall. Four mutually exclusive conditional branches retain the strict scope-wide/namespace and initial/continuation request variants. Tool and property descriptions repeat those combination rules for clients that retain the flat properties but discard composition constraints. A scope-wide request includes compatible legacy and canonical records:

{"scope":"project"}

Supplying namespace_id selects canonical records only. With a namespace, collection selection is omission-sensitive:

{"scope":"project","namespace_id":"mnemosyne"}

Omitted collection_id means collectionless records and every collection in the namespace. Native JSON null means collectionless records only:

{
  "scope": "project",
  "namespace_id": "mnemosyne",
  "collection_id": null
}

A string selects one exact collection:

{
  "scope": "project",
  "namespace_id": "mnemosyne",
  "collection_id": "decisions"
}

The string "null" is the literal collection ID null; clients must send an actual JSON null for collectionless selection. A collection selector without a namespace is invalid. Requests never accept a path, root, filename, query, tag, sort key, arbitrary filter, content field, or cross-scope selector.

Initial requests may include page_size from 1 through 100; the default is 50. When more selected records remain, repeat the exact selector with the returned opaque cursor and omit page_size:

{
  "scope": "project",
  "namespace_id": "mnemosyne",
  "collection_id": "decisions",
  "cursor": "<opaque cursor>"
}

A successful response is always status: ok, including an empty inventory:

{
  "status": "ok",
  "memories": [
    {
      "reference": {
        "schema_version": 2,
        "scope": "project",
        "namespace_id": "mnemosyne",
        "collection_id": "decisions",
        "id": "mem_0123456789abcdef0123456789abcdef"
      },
      "title": "Discovery contract",
      "inspectability": "exact",
      "kind": "decision",
      "lifecycle": {"state": "archived"}
    }
  ],
  "page": {
    "number": 1,
    "count": 1,
    "total_count": 1,
    "total_pages": 1,
    "truncated": false,
    "next_cursor": null
  }
}

Legacy items contain only their version-1 reference, nullable title, and inspectability. Canonical items additionally contain kind and lifecycle state, but not lifecycle revision. Listing never returns content, tags, labels, language, provenance, timestamps, paths, fingerprints, retrieval scores, or match evidence.

Ordering is fixed: schema version first, then public identity. Collectionless canonical records precede collected records within a namespace. Duplicate legacy sources remain separate countable items and use their unexposed relative paths only as a final ordering tie-breaker. Their shared reference is marked inspectability: ambiguous; exact inspection continues to return ambiguous_reference for that reference.

Cursors are authenticated, bound to the exact selector and complete selected valid-record snapshot, and fixed to the original page size. Selected additions, removals, relocations, byte rewrites, lifecycle changes, or valid/invalid transitions return status: conflict, code stale_cursor. Cursors also become stale after a server restart because their key and process marker are not persisted. Malformed, current-process-tampered, selector-mismatched, or otherwise incompatible cursors return status: invalid_request, code invalid_cursor. Start a fresh listing after either cursor error.

Other stable errors are invalid_request, invalid_scope, invalid_namespace, invalid_collection, invalid_page_size, candidate_limit_exceeded, memory_source_unavailable, and internal_error. Candidate overflow and unsafe or unavailable selected sources fail without a partial inventory. Listing does not initialize a missing memory root, create an index or snapshot store, persist cursors, or change any record. Its one terminal log event contains only bounded outcome, selector-presence, count, page, and error metadata; it omits selector values, IDs, titles, cursors, content, paths, fingerprints, exception details, and tracebacks.

Inspecting Memory

memory_inspect is registered by default as a read-only Tool. It accepts exactly one reference object and no path, memory root, filename, query, lifecycle selector, or broad list selector. Use the versioned reference returned by memory_recall or add schema_version: 2 to the canonical identity returned by memory_remember:

{
  "reference": {
    "schema_version": 2,
    "scope": "project",
    "namespace_id": "mnemosyne",
    "collection_id": "decisions",
    "id": "mem_0123456789abcdef0123456789abcdef"
  }
}

A legacy selector is limited to its actual identity:

{
  "reference": {
    "schema_version": 1,
    "scope": "preference",
    "id": "rainy-weekend"
  }
}

Canonical success returns status: ok, the versioned structured reference, and the complete user-visible version-2 record: scope, namespace, optional collection, kind, language, nullable title, content, tags, provenance, lifecycle, and timestamps. Active and archived records use the same result shape; exact inspection does not filter archived memory. Legacy success returns only its versioned reference and fields that version 1 can represent: schema version, ID, nullable title, content, and tags. It does not invent namespace, collection, kind, language, provenance, lifecycle, or timestamps.

Invalid references, missing records, ambiguous legacy IDs, excessive legacy candidates, unsafe or unavailable sources, and unexpected failures return bounded Tool errors with stable codes: invalid_reference, not_found, ambiguous_reference, candidate_limit_exceeded, memory_source_unavailable, or internal_error. Results never include paths, fingerprints, retrieval scores, storage wrappers, or unrelated records.

Inspection never creates, changes, migrates, archives, restores, or deletes a record and does not initialize a missing memory root. Its one terminal log event contains only the outcome, stable code/field where applicable, reference schema version, and scope. It omits IDs, memory text and metadata, complete arguments, paths, exception details, and tracebacks. Shared skipped-record warnings likewise contain only scope and a bounded reason.

Memory scopes, records, paths, storage, retrieval, listing selection, ordering, pagination, cursor policy, content policy, and lifecycle policy live in the plugin-owned mymcp/plugins/mnemosyne/memory/ domain rather than inside individual MCP Tools. The domain includes mutation-disabled revise, archive, restore, and physical-delete primitives. memory_remember, memory_revise, paired memory_archive / memory_restore, and memory_forget are absent from discovery and dispatch unless their independent operator gates enable them. Enabling one gate does not enable another capability.

The MCP host composes kind-qualified ActivatedTool values into PluginContribution objects, validates host-owned public bindings, and builds one immutable HostRuntime. The host binds list_tools to the complete final surface and retains trusted origins, effect/consent metadata, plugin inventory, and an opaque runtime generation. Duplicate plugin, capability, or public names fail composition rather than overwriting a registration.

The explicit production bootstrap reads the fixed packaged inert Mnemosyne manifest, validates exact definition and selected-contribution parity before generation construction, then composes the trusted Mnemosyne 0.1.0 adapter over canonical registrations. The plugin resolves immutable mutation settings once, selects and orders memory Tools, and lazily resolves the configured root for each validated operation. Runtime-bound MCP dispatch, tools/list, and list_tools all use the same runtime registry. mymcp/settings.py retains only host server/process identity; plugin-owned configuration.py retains the unchanged MNEMOSYNE_*, ~/.mnemosyne/config.toml, and memory-root contracts. Handlers remain narrow operation adapters. Synthetic multi-contribution tests prove ordered aggregation, complete reporting, dispatch, and collision rejection. This static bundled-plugin seam preserves Mnemosyne's public server identity, Tool names, schemas, ordering, gates, results, errors, configuration, and storage contracts. The manifest grants no requested authority. This is not dynamic discovery, adds no public Tool namespacing, and provides no installation or lifecycle system.

Revising Memory

Revision is disabled by default and has an independent gate. Persist it in the fixed settings file or use the process override:

[memory]
revise_enabled = true

export MNEMOSYNE_MEMORY_REVISE_ENABLED=true

Only exact lowercase true and false are accepted. Restart the server and reconnect or restart the MCP client after changing enablement. Clients that cannot require approval for every complete exact replacement must leave revision disabled.

Inspect the record immediately before proposing revision. Recall returns an inspect-compatible reference, but inspection supplies the complete current record, lifecycle state, and revision needed for review and approval. Every revision request contains six unconditional fields plus collection_label when the reference identifies an existing collection:

{
  "reference": {
    "schema_version": 2,
    "scope": "preference",
    "namespace_id": "tea",
    "collection_id": "favorites",
    "id": "mem_0123456789abcdef0123456789abcdef"
  },
  "expected_revision": 3,
  "namespace_label": "Tea",
  "collection_label": "Favorites",
  "title": "Japanese green tea",
  "content": "The user enjoys sencha and gyokuro.",
  "tags": ["tea", "japanese-green-tea"]
}

namespace_label and title may be null, but their keys remain required. When reference.collection_id is a string, collection_label is also required and may be either a replacement string or native JSON null. When reference.collection_id is null, the record has no collection and collection_label may be omitted; an explicit native JSON null remains accepted. The literal string "null" is text, not JSON null. Content is nonblank and at most 4,000 characters; tags contain zero to ten unique normalized items of at most 50 characters. Unknown fields, legacy references, paths, patches, language or identity changes, provenance, lifecycle targets, timestamps, and model confirmation/consent fields are rejected.

Clients that stringify structured Tool fields receive the existing one-layer compatibility normalization: a JSON-encoded reference object, tags array, or integer revision is decoded where its schema position disallows strings. A nullable text position permits real strings, so string "null" is deliberately not coerced. Collectionless callers that cannot emit top-level JSON null should omit collection_label.

The caller replaces only namespace label, the label of an existing collection, title, content, and tags. Mnemosyne preserves schema version, record ID, scope, namespace kind/ID, collection presence/ID, memory kind, language, provenance, lifecycle state, and created_at. On change, the server increments lifecycle revision once and updates updated_at. Labels never rename directories or relocate the record. Revision atomically replaces the same source file and creates no second record, patch history, backup, diff, or tombstone.

Success returns only status, canonical versioned reference, and lifecycle:

{
  "status": "revised",
  "reference": {
    "schema_version": 2,
    "scope": "preference",
    "namespace_id": "tea",
    "collection_id": "favorites",
    "id": "mem_0123456789abcdef0123456789abcdef"
  },
  "lifecycle": {"state": "active", "revision": 4}
}

revised means one atomic replacement occurred. A current-revision normalized no-op returns already_current without a write, timestamp update, or revision increment; stale revision conflicts before no-op detection. Archived memory may be revised but remains archived, inspectable, and excluded from normal recall. Active memory remains recall-eligible.

The shared bounded content policy checks every replacement label, title, content value, and tag before storage access. It refuses recognized private-key, credential, authorization-header, JWT-shaped, payment-card, and SSN-shaped values with disallowed_content. Compact-token recognition requires JOSE compact-serialization evidence—a base64url JSON header segment beginning eyJ followed by two further segments—so ordinary dotted text such as versions, dates, hostnames, and module paths is stored normally. This is signature detection, not complete semantic DLP; secrets and sensitive personal data remain prohibited.

Both memory_revise and memory_remember make a content refusal actionable without echoing the rejected value. The Tool error contains the first bounded caller-visible field, one broad reason, and stable remediation guidance:

{
  "status": "refused",
  "code": "disallowed_content",
  "field": "content",
  "reason": "compact_token_shape",
  "message": "memory field resembles content that Mnemosyne does not store; review the named field and retry only if the user confirms that the formatting is benign"
}

The fixed reason vocabulary is private_key_shape, credential_shape, compact_token_shape, payment_card_shape, and government_identifier_shape. Remember maps namespace IDs/labels to field: namespace, collection IDs/labels to field: collection, and reports title, content, or tags directly. Revision reports its flat replacement fields: namespace_label, collection_label, title, content, or tags. Tags never expose an index.

Only the deterministic first match is reported. Results and logs never include the matched value, offset, regular expression, provider-specific detector, fingerprint, or rejected content. The metadata is diagnostic, not permission to evade policy: do not obfuscate suspected sensitive data. Retry only after the user verifies that the formatting is benign and approves the complete revised call.

Stable bounded error codes are invalid_reference, invalid_expected_revision, invalid_record, invalid_collection, disallowed_content, mutation_disabled, not_found, revision_conflict, write_conflict, memory_source_unavailable, replacement_outcome_uncertain, and internal_error. Error statuses are, as applicable, invalid_request, refused, policy_error, not_found, conflict, storage_error, uncertain, or internal_error.

replacement_outcome_uncertain means the atomic replacement may already be visible, but parent-directory durability was not confirmed. Do not retry blindly: inspect the same reference first. If the requested replacement is already present, do not retry; otherwise use the inspected current revision and obtain new exact per-call approval. If inspection is unavailable, the outcome remains uncertain.

Logger mcp.memory_revise emits one terminal event containing only outcome, stable code/field where applicable, schema version, scope, and successful lifecycle state/revision. It omits IDs, labels, title, content, tags, complete arguments, paths, fingerprints, exception details, and tracebacks.

list_tools prefixes the discovered Tool names with the same static server version exposed by MCP initialize and /version, for example Server: mymcp 0.2.0.. Restart the server and reconnect the client after an upgrade; a prior marker identifies a stale process.

Archiving and Restoring Memory

Archive and restore are disabled by default and are enabled together. Persist the operator choice in the same fixed settings file:

[memory]
archive_restore_enabled = true

Or use the process override:

export MNEMOSYNE_MEMORY_ARCHIVE_RESTORE_ENABLED=true

Only exact lowercase true and false are accepted. This gate is independent of remember_enabled; enabling either capability does not enable the other. Restart the server and MCP client after changing enablement.

Each lifecycle request contains exactly the canonical version-2 reference from memory_inspect and its current positive revision:

{
  "reference": {
    "schema_version": 2,
    "scope": "project",
    "namespace_id": "mnemosyne",
    "collection_id": "decisions",
    "id": "mem_0123456789abcdef0123456789abcdef"
  },
  "expected_revision": 1
}

The request accepts no legacy reference, path, root, filename, query, content, target state, timestamp, or consent field. Inspect immediately before proposing the mutation so the complete exact reference and current revision can be shown for approval. A stale revision always returns revision_conflict, even if the record has since reached the requested state.

Archive changes active to archived, increments revision once, updates updated_at, and atomically replaces the same file. Archived memory is excluded from normal recall but remains available through exact inspection. Restore performs the inverse transition and returns the record to recall. A current-revision archive of archived memory returns already_archived; a current-revision restore of active memory returns already_active. Idempotent outcomes do not write, increment revision, or update metadata.

Success contains only status, canonical versioned reference, and lifecycle:

{
  "status": "archived",
  "reference": {
    "schema_version": 2,
    "scope": "project",
    "namespace_id": "mnemosyne",
    "collection_id": "decisions",
    "id": "mem_0123456789abcdef0123456789abcdef"
  },
  "lifecycle": {"state": "archived", "revision": 2}
}

The four normal statuses are archived, already_archived, restored, and already_active. Stable bounded error codes are invalid_reference, invalid_expected_revision, mutation_disabled, not_found, revision_conflict, write_conflict, memory_source_unavailable, replacement_outcome_uncertain, and internal_error. Results never expose record content, paths, fingerprints, or storage wrappers.

For archive and restore, replacement_outcome_uncertain means the new lifecycle state may already be visible but directory durability was not confirmed. Inspect the same reference before any newly approved retry. This is distinct from forget's deletion-specific uncertain outcome.

Each server call logs one terminal event under mcp.memory_archive or mcp.memory_restore. Logs contain only outcome, stable code/field where applicable, schema version, scope, and successful lifecycle state/revision. They omit IDs, namespace/collection identity, labels, title, content, tags, complete arguments, paths, exception details, and tracebacks.

Permanently Forgetting Memory

Forget is disabled by default and has its own gate. It is not enabled by remember or archive/restore. Persist the operator choice in the fixed settings file:

[memory]
forget_enabled = true

Or use the exact process override:

export MNEMOSYNE_MEMORY_FORGET_ENABLED=true

Only lowercase true and false are accepted. Restart the server and MCP client after changing either source. Clients that cannot require approval for every exact call must leave forget disabled.

Forget accepts the same exact canonical version-2 reference and positive expected_revision request shape shown above for lifecycle mutation. It accepts no legacy reference, path, content, fingerprint, target state, timestamp, or confirmation field. The record must already be archived at that current revision. A stale request returns revision_conflict before state eligibility; a current active record returns not_archived unchanged. Inspect immediately before proposing forget so the complete record, archived state, reference, and current revision can be reviewed for per-call approval.

Success physically removes the one source file, leaves organizational directories intact, creates no tombstone, hidden history, backup, retained hash, or deletion manifest, and returns only:

{
  "status": "forgotten",
  "reference": {
    "schema_version": 2,
    "scope": "project",
    "namespace_id": "mnemosyne",
    "collection_id": "decisions",
    "id": "mem_0123456789abcdef0123456789abcdef"
  }
}

A second call returns not_found; Mnemosyne cannot fabricate idempotent proof without a tombstone. Stable bounded codes are invalid_reference, invalid_expected_revision, mutation_disabled, not_archived, not_found, revision_conflict, write_conflict, memory_source_unavailable, deletion_outcome_uncertain, and internal_error.

deletion_outcome_uncertain means unlink may have succeeded but directory durability was not confirmed. Do not retry blindly. Inspect the same reference: if it is absent, do not retry; if it is present, review it again, require it to remain archived, use its current revision, and obtain new per-call approval. If inspection is unavailable, the outcome remains uncertain.

Forgetting is irreversible within Mnemosyne: restore cannot recover a forgotten record. Physical source deletion is not secure erasure and does not remove filesystem journal data, snapshots, backups, MCP-client history, logs outside Mnemosyne, or external copies. Logger mcp.memory_forget emits one terminal content-free event with only bounded outcome/code/field, schema version, and scope metadata.

Remembering Memory

Remember is disabled by default. With no setting, an accepted disabled setting, or the exact environment value false, tools/list and list_tools omit memory_remember and direct dispatch treats it as unknown.

Persist the operator's choice in the one fixed local settings file:

~/.mnemosyne/config.toml

[memory]
remember_enabled = true

The file is optional and read-only to Mnemosyne. Startup does not create or edit it or its parent. A missing file, empty document, absent or empty [memory] table, absent key, or TOML boolean false all mean disabled. The document may contain only the optional [memory] table, which may contain only the optional TOML booleans remember_enabled, archive_restore_enabled, forget_enabled, and revise_enabled; strings such as "true", unknown keys/tables, and malformed TOML fail startup closed.

For one process-level override, use:

export MNEMOSYNE_MEMORY_REMEMBER_ENABLED=true

Each environment variable overrides only its matching setting. Mnemosyne parses all four mutation environment values first; an invalid supplied value fails startup before file access without being echoed. If any setting remains unresolved, Mnemosyne reads the strict file at most once for unresolved values. The file is bypassed only when all four environment variables are supplied. Every unresolved setting finally defaults to false.

The server reads enablement once at startup, so restart it after changing either source. Restart or reconnect the MCP client as well so it refreshes Tool discovery. Enabling the server-side Tool does not establish user consent. A compatible MCP client must separately show the complete arguments and require user approval for every exact call.

The settings source is limited to 16 KiB of UTF-8 TOML. Mnemosyne rejects a symlinked .mnemosyne directory or file, non-directory/non-regular source, unreadable source, and—on POSIX—a directory or file writable by group or others. Use mode 0700 for ~/.mnemosyne and 0600 for config.toml; non-writable 0755/0644 modes are also accepted because this file is not a secret store. Configuration failures use stable bounded messages and do not expose supplied values, file contents, parser details, underlying exception text, or the absolute settings path.

Nine caller-owned fields are always required, including nullable values. The Tool schema publishes those nine plus optional occurred_at in top-level properties, while retaining the nine-field top-level required list, so clients that keep only a flat object schema can construct both ordinary memories and events. Six scope-specific oneOf branches additionally narrow namespace kind and memory kind for clients that support composition:

{
  "scope": "project",
  "namespace": {
    "kind": "project",
    "id": "mnemosyne",
    "label": "Mnemosyne"
  },
  "collection": {
    "id": "decisions",
    "label": "Decisions"
  },
  "kind": "decision",
  "language": "en",
  "title": "Remember consent boundary",
  "content": "Durable memory requires approval for each exact Tool call.",
  "tags": ["architecture", "consent"],
  "origin": "user_approved_proposal"
}

For a completed project occurrence, use kind: "event" and include its known occurrence instant separately from persistence time:

{
  "scope": "project",
  "namespace": {"kind": "project", "id": "mnemosyne", "label": "Mnemosyne"},
  "collection": {"id": "events", "label": "Events"},
  "kind": "event",
  "language": "en",
  "title": "Track activated",
  "content": "Track 021 moved to active execution.",
  "tags": ["track-021"],
  "origin": "explicit_user_statement",
  "occurred_at": "2026-07-20T15:00:00Z"
}

occurred_at is required exactly for event, rejected for every other kind, and uses strict UTC-second form YYYY-MM-DDTHH:MM:SSZ. It is immutable after creation: revision has no occurrence replacement field, and archive/restore preserve it. Existing non-event version-2 records remain valid without the field and serialize without an invented null value.

namespace.label, collection, collection.label, language, and title may be null, but their keys remain required. Tags are a required zero-to-ten item array. Namespace/collection IDs are 1–64 safe identifier characters, labels are at most 100 characters, language tags at most 35, title at most 200, content at most 4,000, and each tag at most 50. Unknown fields are rejected at every level. The allowed dimensions are derived from the shared domain:

Scope	Namespace kinds	Memory kinds
self	aspect	attribute
relationship	person, group, relationship	perspective, summary
preference	domain	preference
practice	domain	practice
project	project	decision, constraint, state, event, question, reference, summary
knowledge	topic	reference, summary

The shared domain defines writing guidance for every allowed (scope, kind) pair. Complete scope branches publish only their applicable guidance, while the top-level kind description groups all guidance by scope for clients that discard schema composition. Shared kinds such as summary and reference therefore retain scope-specific meaning.

Public origin is caller-supplied provenance context: use explicit_user_statement for a direct user statement or user_approved_proposal for an approved proposed memory. Origin is not consent; consent remains the MCP client's approval of the complete exact call. recorded_via is instead server-assigned. Callers cannot supply a filesystem path, record ID, persistence timestamps, lifecycle state, revision, recorded_via, or model-authored confirmation/consent field. The event-only occurred_at value is the sole caller-owned structural time.

A new memory returns only its status, structured reference, and lifecycle:

{
  "status": "remembered",
  "reference": {
    "scope": "project",
    "namespace_id": "mnemosyne",
    "collection_id": "decisions",
    "id": "mem_0123456789abcdef0123456789abcdef"
  },
  "lifecycle": {
    "state": "active",
    "revision": 1
  }
}

An exact active duplicate returns already_exists; an archived duplicate returns existing_archived. Both identify the existing reference/lifecycle and write no second file. Validation, content refusal, disabled mutation, candidate overflow, generated-ID conflict, storage failure, and unexpected failure return bounded Tool errors with stable codes and no path or submitted content. Content refusal uses the bounded field/reason contract documented under revision above. Validation uses invalid_scope, invalid_namespace, invalid_collection, invalid_kind, invalid_record, or invalid_origin; policy refusal uses disallowed_content; disabled mutation uses mutation_disabled; bounded discovery uses candidate_limit_exceeded; publication conflict uses write_conflict; storage/path failures use memory_source_unavailable; and an unexpected failure uses internal_error.

Event duplicate identity includes occurred_at: otherwise-equal events at the same occurrence instant are duplicates, while events at different instants are distinct. Existing non-event duplicate identity is unchanged.

Before duplicate discovery or directory creation, the shared policy rejects recognized private-key blocks, declared provider token/API-key prefixes, Basic/Bearer authorization headers, credential assignments and credential-bearing URLs, compact JWT-shaped values, Luhn-valid 13–19 digit payment-card values, and SSN-shaped identifiers. It inspects namespace/collection IDs and labels, title, content, and tags. This is intentionally bounded signature detection, not semantic DLP: false positives and false negatives remain possible. Unrecognized secrets and sensitive personal facts remain prohibited, and the user must review the exact proposed arguments before approval.

Remember logs one terminal event containing only outcome, stable error code/field where applicable, scope, namespace kind, memory kind, collection-present state, origin, and generated ID/lifecycle after a domain result. Logs never include labels, title, content, tags, language, occurred_at, complete arguments, rejected values, filesystem paths, exception messages, or tracebacks.

Filesystem Memory

The default memory root is:

~/.mnemosyne/memory

No manual first-run initialization is required. When an enabled, valid, policy-accepted memory_remember call reaches canonical creation, Mnemosyne lazily creates any missing directories from the nearest existing ancestor through the record directory. Every newly created directory uses private mode 0700 on POSIX, and the record file uses mode 0600. Existing directories are left unchanged rather than chmodded.

Resolving either settings source, starting the server, recalling memory, disabled or invalid remember calls, and content-policy refusals do not initialize the memory root. Filesystem initialization therefore remains behind the existing operator-enable and per-call consent boundaries.

Set an explicit root for another local location:

export MNEMOSYNE_MEMORY_ROOT=/path/to/memory

The memory domain recognizes only the six fixed scope directories beneath the memory root:

memory/
  self/
  relationship/
  preference/
  practice/
  project/
  knowledge/

Legacy version-1 records remain readable in their existing locations. Their minimal format is:

{
  "schema_version": 1,
  "id": "rainy-weekend",
  "title": "Rainy weekend activities",
  "content": "On rainy weekend afternoons, the user prefers museums or quiet cafés.",
  "tags": ["leisure", "rainy-day", "weekend"]
}

Version-1 schema_version, id, and content are required; title and tags are optional. No background migration or rewrite occurs.

Canonical version-2 records are self-contained and use a deterministic path:

<scope>/<namespace-id>/<collection-id?>/<memory-id>.json

For example:

project/mnemosyne/decisions/mem_0123456789abcdef0123456789abcdef.json

{
  "schema_version": 2,
  "id": "mem_0123456789abcdef0123456789abcdef",
  "scope": "project",
  "namespace": {
    "kind": "project",
    "id": "mnemosyne",
    "label": "Mnemosyne"
  },
  "collection": {
    "id": "decisions",
    "label": "Decisions"
  },
  "kind": "decision",
  "language": "en",
  "title": "Shared memory ownership",
  "content": "Canonical memory concepts belong to the shared memory domain.",
  "tags": ["architecture", "memory-domain"],
  "provenance": {
    "origin": "explicit_user_statement",
    "recorded_via": "memory_remember"
  },
  "lifecycle": {
    "state": "active",
    "revision": 1
  },
  "created_at": "2026-07-18T12:00:00Z",
  "updated_at": "2026-07-18T12:00:00Z"
}

Scope, namespace, optional collection, kind, language, content, tags, provenance, persistence timestamps, optional event occurrence time, and lifecycle are separate dimensions. IDs determine paths; mutable labels do not. Record metadata must agree with its location. Unknown fields or mismatched paths make a version-2 record invalid.

Canonical project events persist "kind": "event" and a strict "occurred_at" value. Exact inspection returns that field; recall and listing retain their existing relevance and identity ordering rather than becoming chronological views. Mnemosyne does not create timeline resources, infer causality, automatically supersede state, or make event records fully append-only. A first-class many-to-many temporal model remains deferred until a concrete usage need demonstrates it.

Version-2 records are either active or archived; normal recall excludes archived records. Revision replaces the same file atomically without retaining hidden prior content. Forgetting is physical deletion with no tombstone. Remember creation, complete-state revision, reversible archive/restore, and archived-only physical forget are exposed through explicit MCP Tools behind independent default-off gates. Revision atomically replaces the same canonical file and retains no hidden prior content.

All record files are limited to 64 KiB, content to 4,000 characters, and titles to 200 characters. Invalid, oversized, too-deep, or unsafe records are skipped and logged without their content.

Recall retrieval case-folds and tokenizes the query, relative path, title, content, and record tags. Exact request-tag overlap has the strongest weight, followed by title, path/record-tag, and content matches. Ties are resolved deterministically. Symlinks are rejected, no more than 1,000 candidate files are accepted in one scope, and no more than five records are returned. Files remain the source of truth: inspect them directly and delete a record by deleting its file.

There is no required manifest or persistent content-bearing index. Exact inspection and complete listing use structured selectors and never accept a filesystem path. Scope-wide listing applies the same 1,000-candidate bound; canonical namespace and collection selectors narrow to their deterministic safe container before candidate counting. Collectionless and exact-collection discovery scan only the canonical path level. Listing fails instead of returning a partial inventory when the selected container exceeds the bound. Every mutation remains disabled unless its operator gate enables it and the MCP client can require approval for every exact call. A model-provided confirmation field is not consent.

MCP Validation

An MCP request envelope must be an object. Otherwise the server returns JSON-RPC error -32600 with the message Invalid Request and id: null. When present, its params value must also be an object. Otherwise the server returns JSON-RPC error -32602 with the message Invalid params and preserves the request ID. For tools/call, params.arguments must likewise be an object when present; non-object Tool arguments receive the same -32602 response before Tool selection, schema-aware normalization, or handler dispatch.

MCP notifications omit id and receive HTTP 202 with no JSON-RPC response body. notifications/initialized and notifications/cancelled are accepted; cancellation is currently a no-op because tool handlers complete synchronously.

For compatibility with MCP clients that stringify structured Tool arguments, dispatch uses the selected Tool's advertised input schema to remove at most one extra JSON-encoding layer. Normalization occurs only where the schema disallows a string and the decoded JSON value has a permitted type. Correctly typed values are unchanged; fields that permit strings are never decoded; malformed, wrong-type, or repeatedly encoded values continue to the Tool's normal bounded validation. The advertised Tool schemas and canonical memory formats remain unchanged. For memory_list, a stringified page size can therefore normalize to an integer, while collection_id strings—including "null"—remain strings because that field legitimately permits them. Collectionless selection requires native JSON null.

Intended Roles

MyMCP is intended to become a local-first MCP host and client-neutral governance gateway. It owns Tool composition, identity, plugin contracts, routing, and reusable host mechanisms without absorbing integration-specific domain policy.

Its approved Tools-only target separates one host-owned logical plugin API from concrete bundled implementations and isolated external execution adapters. It retains qualified (plugin_id, capability_kind, capability_local_id) origin internally and binds that identity to the endpoint's flat MCP Tool name. Host bootstrap remains explicit. Native installation, lifecycle generations, isolation, authenticated client policy, host-verifiable exact-call approval, and bounded security audit remain unimplemented.

Mnemosyne remains the bundled user-governed memory domain. It gives agents controlled access to approved durable records and bounded retrieval while permanently preserving its Tool, configuration, storage, record, logging, consent, and memory-domain identity. Its configuration, memory domain, MCP adapters, and plugin composition live under mymcp/plugins/mnemosyne/.

Non-Goals

MyMCP and its built-in Mnemosyne domain are not intended to provide:

• generic shell execution or unrestricted filesystem access;
• secret storage;
• hidden mutation or governance that bypasses visible user consent;
• a multi-user platform before an explicit threat model supports one;
• MCPB as the native MyMCP plugin format; or
• unknown-code loading, lifecycle, marketplace, or isolation claims before artifact, supervision, authority, failure, and approval contracts are implemented and validated.

Running the Server

Clone the canonical MyMCP repository target:

git clone https://github.com/flyset/MyMCP.git
cd MyMCP

The canonical repository is https://github.com/flyset/MyMCP; the former URL redirects here.

The MCP endpoint is exposed at:

http://127.0.0.1:8000/mcp

Additional operational endpoints:

• http://127.0.0.1:8000/health
• http://127.0.0.1:8000/version

Create a local virtual environment and install the project in editable mode:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .

Start the development server with auto-reload:

mymcp-dev

Start the server without auto-reload:

mymcp

The supported direct ASGI target is the application factory mymcp.app:create_production_app with Uvicorn factory mode. The former internal module-level target mymcp.app:app no longer exists.

Run the test suite after installing the test extra:

mymcp-test

OpenCode Configuration

The tracked opencode.json registers this server as remote connection and agent mymcp. At both policy levels it denies the broad mymcp_* prefix first, then allows exactly four read-only Tools (list_tools, memory_recall, memory_inspect, and memory_list), then asks for the five exact mutation Tools (remember, revise, archive, restore, and forget). The order matters: OpenCode uses the last matching Tool-name rule and per-agent permissions can override top-level rules. The tracked and ignored OpenCode migration is complete; the legacy Claude configuration is excluded. The normal MyMCP endpoint and final client reconnect were verified. In isolated checks, one approved mutation was approved once and one rejected mutation made no Tools call; neither check changed the memory root.

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "mymcp_*": "deny",
    "mymcp_list_tools": "allow",
    "mymcp_memory_recall": "allow",
    "mymcp_memory_inspect": "allow",
    "mymcp_memory_list": "allow",
    "mymcp_memory_remember": "ask",
    "mymcp_memory_revise": "ask",
    "mymcp_memory_archive": "ask",
    "mymcp_memory_restore": "ask",
    "mymcp_memory_forget": "ask"
  },
  "mcp": {
    "mymcp": {
      "type": "remote",
      "url": "http://127.0.0.1:8000/mcp",
      "enabled": true
    }
  },
  "agent": {
    "mymcp": {
      "permission": {
        "mymcp_*": "deny",
        "mymcp_list_tools": "allow",
        "mymcp_memory_recall": "allow",
        "mymcp_memory_inspect": "allow",
        "mymcp_memory_list": "allow",
        "mymcp_memory_remember": "ask",
        "mymcp_memory_revise": "ask",
        "mymcp_memory_archive": "ask",
        "mymcp_memory_restore": "ask",
        "mymcp_memory_forget": "ask"
      }
    }
  }
}

This client permission does not enable mutation on the server. The operator must separately set the applicable remember, revise, archive/restore, or forget server gate and restart the server. After changing opencode.json or agent policy, quit and restart OpenCode so it reloads configuration and Tool discovery.

For every proposed mutation call, review the complete Tool arguments and choose once. Rejecting a mutation prevents the request from reaching Mnemosyne and therefore produces no write. Do not start OpenCode with --auto or enable interactive auto-approval while memory mutation is enabled: each bypasses approval on later exact calls. If either mode was used, disable server mutation and restart both the server and OpenCode before continuing. Any additional per-agent permission override must preserve this order: broad mymcp_* denial first, explicit lower-breadth read-only allows (including memory_list) next, and the exact mutation asks last.

Integrating Mnemosyne with a Repository

After Mnemosyne's memory Tools are available to an AI client, add a repository-root MEMORY.md. It is the repository-local operating contract: it tells agents which project context to retrieve, where durable records belong, and when records should be created, revised, archived, or forgotten.

Keep the repository, Git history, backlog, and documentation as the authoritative engineering record. Mnemosyne preserves concise, user-approved orientation, rationale, and delivery history across sessions.

Start with docs/GETTING_STARTED.md for a project memory-map template and integration guidance. Agents should read docs/MANUAL.md for general safe Tool-use guidance before operating on memory.

Roadmap Shape

The roadmap converges on the complete approved end state rather than freezing a bundled-only callable or deferring identity and security boundaries:

1. Delivered: kind-qualified Tools-only identity, effect/consent metadata, host-owned bindings, immutable HostRuntime, a trusted Mnemosyne adapter, explicit bootstrap, runtime injection, immutable definition contracts, strict 64 KiB manifest parsing, complete packaged declaration, exact parity, static validation before generation, wheel inclusion, and the one canonical Mnemosyne implementation under mymcp/plugins/mnemosyne/—without changing public behavior.
2. Released: MyMCP/mymcp 0.2.0 host, application, package, repository, and tracked-client identity preserve all Mnemosyne plugin/domain identities. The public non-draft, non-prerelease release is tagged mymcp-v0.2.0 at c2852bc.
3. Add side-effect-free native wheel inspection, digest-bound receipts, immutable managed environments, bounded configuration/secret references, and persisted bindings without granting execution authority.
4. Add isolated external activation only after a local unknown-code threat model, supervision, killability, resource bounds, and default-deny filesystem/network enforcement are validated.
5. Add generation lifecycle, required/optional failures, drain, quarantine, update/rollback and data migration, removal/preservation, and separate purge.
6. Build authenticated local-client routing, policy-filtered discovery/dispatch, single-use exact-call approval, and bounded security audit behind the one machine-local endpoint.
7. Generalize host services only after a second real plugin proves them reusable.

MCPB may later package MyMCP as one complete server. It is not the native MyMCP plugin distribution, manifest, or execution boundary.

See docs/PLUGIN_ARCHITECTURE.md for the target and incremental migration rules.

See VISION.md for MyMCP's broader scope and boundaries. See docs/MNEMOSYNE_VISION.md for the built-in memory domain's notebook model, safety contract, and direction.

See docs/ARCHITECTURE.md for the current code organization.

See docs/PLUGIN_ARCHITECTURE.md for the approved plugin end state and roadmap alignment.

See docs/GETTING_STARTED.md for repository-level project-memory integration, and docs/MANUAL.md for operating guidance for agents using memory Tools.

See docs/AI_WORKFLOW.md for contribution and verification gates, and docs/GLOSSARY.md for canonical terms and public-contract language.