Automatic CodeGraph builds: Trusted bootstrap starts a project's first
build in a background process. Default and per-project switches persist;
disabling automation preserves the graph. Supported file-write events
trigger incremental refresh.
Clear graph navigation: Switch between a file overview and symbols,
filter by file, inspect neighboring relations, and use zoom controls.
Local-first MCP memory governance for coding agents — Codex, Claude Code, Cursor, Grok, and Trae. Shared rules, deduplication, token insights, audit, and rollback.
2026-10-28
ดาว
4
เครื่องมือ
—
อัปไทม์
—
Bounded temporary storage: Ordinary graph queries read SQLite directly
instead of retaining full-database copies. Validation and installation
temporary files use the owning project/deployment cache and are cleaned up
when the operation ends.
v0.7.13 strengthens mandatory-rule replacement and recovery, improves
audience-aware identity handling, and improves conflict resolution:
Safe mandatory-rule replacement: Replacements validate the final
mandatory package against audience matching, canonical rules, deduplication,
sensitivity checks, and publication budgets. Equivalent unlocked predecessors
retire in the same transaction; failed validation rolls back the complete
update. See the replacement and recovery details.
Native Agent/group audience matching: Matching and deduplication no longer
split the same native audience solely by provider or runtime role. Distinct
Agent identities remain separate, and project-scoped audiences retain their
project boundaries. Provider repair uses the verified identity for its target
provider.
Readable, atomic conflict resolution: Conflict views preserve readable
peer information, and resolution keeps peer groups intact while applying the
selected changes atomically. Ambiguous conflicts remain unresolved.
Cursor Hook protection: Cursor MemoryGuard Hooks now use a 30-second
timeout, up from 15 seconds, with failClosed: true.
v0.7.14 only corrects release metadata for the GitHub repository rename; it does not change runtime behavior. See the v0.7.14 release note.
Usage events distinguish measured_cached_input from
measured_cache_write_input. measured_cache_coverage.cache_read and
cache_write report complete, partial, or unavailable; measured zero
remains 0, while missing provider data remains None/unavailable.
Character-based estimates remain explicitly labelled
estimated mg_deterministic_unit, never provider tokens.
Run the benchmark only against an authorized local workspace:
Read the benchmark guide for measured,
estimated, derived, and unsupported semantics. Use the demo recording
checklist for a sanitized walkthrough. The
repository's synthetic graph artwork is not a live product capture; it is not
evidence of usage or savings.
Major V2 refactor in v0.6.0
v0.6.0 was a production data-plane refactor, not a storage-only upgrade:
Authoritative V2 domains: Memory, Rules, Evidence, Content, Runtime, Projection, Assets, CodeGraph, Skills, and System state are separated into explicit SQLite domains with governed boundaries.
Explicit cutover:V1_ACTIVE → V2_BUILDING → V2_READY → V2_ACTIVE is fail-closed; V2 never silently falls back to legacy stores or dual-writes after READY/ACTIVE.
Lossless migration: frozen-source preparation uses coherent SQLite online backups, validates source/target evidence, rechecks live-source drift, and preserves V1 data plus migration backups for rollback.
Native routing: MCP, CLI, GUI, and Hook surfaces are classified explicitly; the release closed the 233-surface cutover with 138 implemented routes, 95 retired routes, and zero neutral/blocker routes.
Governed intelligence: Rule lifecycle and RuleMerge, extraction/enrichment, External MCP import, provider control-plane, conversation history, Knowledge Library, and GUI governance all use the V2 evidence and decision paths.
Operational evidence: Reference Audit, per-domain SQLite health, guarded maintenance, rollback evidence, and safe unbound diagnostics are part of readiness and operations.
Why MemoryGuard
Persistent memory solves storage. It does not solve governance.
When several coding agents write into the same context, records become
duplicated, stale, contradictory, over-broad, or unsafe to reuse. MemoryGuard
sits between coding agents and their shared memory to keep that context usable.
Without governance
With MemoryGuard
Notes accumulate without a canonical state
Writes are classified, deduplicated, superseded, or surfaced as conflicts
A correction silently destroys the old value
Evidence and supersede chains preserve what changed and why
Tokens and credentials can remain active
Sensitive-looking content is quarantined from active memory
Every write needs manual approval
Agents write normally; people review exceptions and outcomes
Raw chat logs leak into future context
Conversation history remains a separate, explicitly read evidence archive
This package exposes a local stdio MCP server as io.github.MerakOsiris/memoryguard.
Registry metadata is kept in server.json, and the marker above
ships with the PyPI package README. Releases are published through GitHub OIDC
to PyPI and the official MCP Registry. Verify the current package version and
the Registry entry's active/latest state through their live public records.
1. Install
python -m pip install agent-memguard
For the desktop governance console:
python -m pip install "agent-memguard[gui]"
2. Authorize the current project
memoryguard source add .
3. Connect or repair your coding agent
Global provider configuration is rebuilt from the real binding in the canonical user data home. The command is idempotent and removes superseded MemoryGuard project-level overrides after a successful global takeover.
# Repair one provider
memoryguard provider repair claude
memoryguard provider repair codex
memoryguard provider repair cursor
memoryguard provider repair trae
# Repair every detected provider
memoryguard provider repair all
Restart the host after installation, then verify the integration:
memoryguard doctor
memoryguard mcp-status
memoryguard hooks status --provider all
Launch the desktop console:
memoryguard gui
memoryguard-gui . remains available for desktop shortcuts. A bare
memoryguard gui always opens the canonical user-level control directory
(default %LOCALAPPDATA%\MemoryGuard on Windows), so running it from a project
or from C:\Windows\System32 cannot silently switch databases.
MEMORYGUARD_WORKSPACE is an explicit operator override; an explicit
memoryguard gui <project-path> or memoryguard gui --workspace <project-path>
selects a specific workspace.
It does not remember a previously selected project or open a folder picker.
On Windows, memoryguard gui detaches the native window from the terminal, so
closing PowerShell does not close the GUI.
Codex/Router binds MemoryGuard to the stable local Codex program and control
installation. An account profile is an endpoint/alias, not a new memory owner:
switching profiles automatically discovers or repairs the profile and reuses the
verified Agent binding and active group. Request identity remains fail-closed;
this does not share records across machines or with arbitrary accounts.
Upgrade
MemoryGuard currently upgrades through Python's package manager:
python -m pip install --upgrade agent-memguard
memoryguard --version
memoryguard doctor
If you installed the GUI extra, keep it during the upgrade:
There is no package self-update command. The package manager is the
authoritative package-upgrade path; memoryguard upgrade below is the explicit
workspace migration flow, not a package updater.
Upgrade an existing V1 data home
Upgrade the package, then run the verified migration. No workspace, data-home,
apply, or confirmation arguments are required for the normal user-level data
home:
python -m pip install --upgrade agent-memguard
memoryguard --version # confirms installed version
memoryguard upgrade
memoryguard doctor
The command prepares V2, validates the frozen and live source evidence,
migrates Agent/Group control, activates only after all gates pass, and removes
only the backup batch belonging to that successful migration. Re-running it on
V2_ACTIVE is idempotent. For a zero-write report, use:
memoryguard upgrade --preview
Advanced explicit workspace/data-home options remain available for operators
managing an isolated installation. A failed gate stays non-active and preserves
its evidence; successful activation does not keep a redundant migration backup.
Existing pre-V2 workspaces: explicit V2 cutover
v0.6.0 never auto-activates an existing workspace. Upgrade the package first,
then use the packaged operator CLI:
# Read-only manifest status
memoryguard-v2 status -w .
# Build a frozen-source V2 shadow and stop at V2_READY
memoryguard-v2 prepare -w . --apply
# Activate only after the prepare result is V2_READY / ready=true
memoryguard-v2 activate -w . --confirm V2_ACTIVE
The prepare step uses coherent SQLite online backups, preserves V1 and
migration-backups, and rechecks live-source drift before READY. Activation
performs another fresh drift check before changing the manifest. Do not delete
legacy V1 data or migration backups as part of the upgrade.
Knowledge Library
The desktop console can turn a selected folder or file set into one governed
local knowledge library. Source files remain where they are; MemoryGuard stores
the searchable index in its user data home instead of copying a runtime
database into every source project. Knowledge metadata never becomes a second
source-body store.
Capability
Current behavior
File/folder ingestion
Add a folder as a book or selected files as documents
Structure
Parse documents, preserve chapter/section context, and create traceable chunks
Retrieval
Full-text search, optional embeddings, and a layered knowledge graph
Natural synchronization
Re-ingest changed files; a partial or failed scan does not silently remove previously indexed content
Lifecycle
Move a book to the library trash, restore it, or explicitly purge its recovery snapshot
Memory candidates
Preview evidence-backed candidates before accepting them into governed long-term memory
Open the desktop console and choose Knowledge Library. Remote embedding or
model-backed indexing is opt-in and requires explicit authorization; local
full-text retrieval remains available without sending source text to a remote
provider. Background imports, re-ingests, and smart rebuilds have durable task
receipts: retrying the same request reuses its task, while reusing that key for
a different request is rejected. A live task for a different request reports
busy rather than claiming that work was accepted.
CodeGraph refresh
The first CodeGraph build is an explicit, confirmed full build. After a scope
has been built, each successful trusted file write can trigger an incremental
refresh for that scope, subject to strict source-path and active-binding
validation. Unchanged content hashes are a no-op; deleted files are retired;
the next context receives one bounded affected receipt. MemoryGuard does not
run a daemon or watcher for this path and does not infer paths from shell or
free-form text. A projectless MCP caller first builds an already-bound directory
source, then passes its codegraph_source_id to select that exact scope for
query, status, update, and graph reads.
Desktop console surfaces
The GUI has eight visible navigation entries: seven governance pages plus a
separate Token usage-and-savings view:
Governance Overview
Data Sources & Agents
Memory Core
CodeGraph
Rules & Habits
Conversation History
Risk Signals & Governance Console
Token Usage & Savings (separate from the seven governance pages)
Agent lists use readable program/provider names; the underlying ID remains
available in the detail view. Empty data is shown as an explicit empty state.
Write and governance lifecycle
%%{init: {"theme":"base","themeVariables":{"background":"#071521","fontFamily":"Arial, sans-serif","fontSize":"14px","primaryTextColor":"#EEF4F8","lineColor":"#557287","edgeLabelBackground":"#071521","clusterBkg":"#0A1A29","clusterBorder":"#27445A"},"flowchart":{"htmlLabels":true,"curve":"basis","nodeSpacing":30,"rankSpacing":42,"padding":14}}}%%
flowchart TD
subgraph Intake["01 · INTAKE "]
direction LR
Write(["Memory write "]):::entry
Scope["Resolve identity<br/>scope · audience "]:::core
Validate{"Authorized? "}:::decision
Reject["Reject<br/>no persistence "]:::danger
Write --> Scope --> Validate
Validate -- NO --> Reject
end
subgraph Organize["02 · ORGANIZE "]
direction TB
Secret{"Sensitive? "}:::decision
Quarantine["Quarantine<br/>outside active set "]:::danger
Compare["Classify · compare<br/>governed records "]:::active
Relation{"Relationship "}:::decision
New["NEW<br/>create active record "]:::result
Duplicate["DUPLICATE<br/>merge provenance "]:::result
Correction["CORRECTION<br/>supersede old record "]:::rule
Conflict["CONFLICT<br/>preserve both sides "]:::danger
Secret -- YES --> Quarantine
Secret -- NO --> Compare --> Relation
Relation --> New
Relation --> Duplicate
Relation --> Correction
Relation --> Conflict
end
subgraph Govern["03 · GOVERN "]
direction LR
Receipt[("Evidence event<br/>version receipt ")]:::store
Review["CLI or desktop review "]:::surface
Action["Correct · merge<br/>restore · delete "]:::rule
Snapshot["Reversible<br/>snapshot "]:::active
Receipt --> Review --> Action --> Snapshot
end
Validate -- YES --> Secret
Quarantine --> Receipt
New --> Receipt
Duplicate --> Receipt
Correction --> Receipt
Conflict --> Receipt
classDef entry fill:#EEF4F8,stroke:#38D5C8,color:#071521,stroke-width:2.4px;
classDef core fill:#12243A,stroke:#557287,color:#EEF4F8,stroke-width:1.5px;
classDef decision fill:#0D3338,stroke:#38D5C8,color:#EEF4F8,stroke-width:2px;
classDef active fill:#0D383A,stroke:#38D5C8,color:#EEF4F8,stroke-width:2px;
classDef result fill:#12243A,stroke:#38D5C8,color:#EEF4F8,stroke-width:1.6px;
classDef rule fill:#3B2C18,stroke:#F3B562,color:#EEF4F8,stroke-width:1.8px;
classDef danger fill:#3A2028,stroke:#EA6A6A,color:#EEF4F8,stroke-width:1.8px;
classDef store fill:#0B1624,stroke:#7F96A8,color:#EEF4F8,stroke-width:1.4px;
classDef surface fill:#EEF4F8,stroke:#38D5C8,color:#071521,stroke-width:2px;
style Intake fill:#081827,stroke:#27445A,stroke-width:1px,color:#EEF4F8
style Organize fill:#081827,stroke:#27445A,stroke-width:1px,color:#EEF4F8
style Govern fill:#081827,stroke:#27445A,stroke-width:1px,color:#EEF4F8
linkStyle default stroke:#557287,stroke-width:1.4px;
The console is not an approval queue. Agents keep moving. MemoryGuard records
the outcome and exposes the evidence needed to correct it later.
What you can govern
Signal
Governance action
Duplicate or stale memory
Inspect the canonical record and supersede chain; restore an earlier version when needed
Conflicting memories
Keep both visible until the conflict is resolved deliberately
Secrets, tokens, or credentials
Quarantine the record so it cannot enter active shared memory
Incorrect automatic organization
Correct, merge, lock, restore, or roll back with evidence
Multiple coding agents
Bind agents to one shared group while preserving source identity and scope
Mandatory rules
Assign rules to an Agent, project, provider, runtime role, or shared group
Rules and history stay separate
MemoryGuard deliberately keeps governed long-term memory and raw conversation
history on different paths.
Surface
Purpose
Context behavior
Rules and habits
Preferences, procedures, corrections, facts, projects, and scoped mandatory rules
Mandatory rules use an independent char/token budget after scope, exclude, conflict, and semantic dedup. Effective count above 20 is a health warning, not a hard block; storage is not capped by count. Sensitive, corrupt, per-item oversize, and aggregate overflow still fail closed with no silent truncation. Ordinary records are recalled when relevant
Conversation history
Local raw-evidence archive with owner and shared-group access controls
Never enters bootstrap automatically; raw text is read only through explicit history tools
Neuron graph
Navigation and governance over memory, rules, projects, agents, and sessions
History nodes contain safe metadata and summaries, not raw chat content
History retrieval is progressive: search results, then a bounded timeline, then
an explicitly selected turn or session. Extracting from history creates a
preview first; it does not silently write a long-term memory.
Global MCP binding, redirect rules, user-level lifecycle Hook
Verified takeover path
Codex
Global MCP binding, redirect rules, user-level lifecycle Hook
Verified takeover path
Cursor
Global MCP binding, redirect rules, user-level lifecycle Hook
Verified takeover path
TRAE
MCP binding and redirect rules
No verified Hook seam; reported as a fallback instead of full takeover
Provider status is reported honestly as redirected, observed, operational, or
unsupported. MemoryGuard does not claim it can disable every host's native
memory when the host exposes no reliable integration point.
V2 uses separate authoritative SQLite domains rather than one shared-memory
database. The runtime reads and writes V2 only after the manifest reaches
V2_ACTIVE; V2_BUILDING and V2_READY never silently fall back or dual-write.
Evidence remains traceable without being treated as automatically trusted memory.
Privacy and safety
MemoryGuard runs as a local MCP stdio server.
All governed data stays local unless you explicitly authorize a remote model
or embedding operation. Optional usage telemetry is local-only: its measured
host token events and deterministic conversion events are stored under
.memoryguard/usage_telemetry.sqlite; it does not upload data. Token savings
are estimates based on MemoryGuard deterministic units, not a provider billing
statement. Hosts without token reporting remain unsupported in the measured
columns.
The Knowledge Library database uses MEMORYGUARD_HOME or the platform user
data directory, so a selected source folder does not receive its own
knowledge database.
V2 authoritative workspace state is separated under .memoryguard/ into
explicit Memory, Rules, Evidence, Content, Runtime, Projection, Assets,
CodeGraph, Skills, and System domains; History, Source, Binding, and Group
control are V2-native surfaces. Legacy V1 artifacts are preserved as local
rollback/audit evidence after cutover and are no longer the active V2 runtime
write path; only memoryguard.migration may read them.
Source scanning is read-only by default.
Mutating governance paths use validation, explicit scope, provenance, and
reversible state.
Quarantined records stay outside active shared memory.
Raw conversation history is never injected into bootstrap automatically.
Shared-group history access follows current active membership and does not
grant deletion rights over another Agent's source.
CLI
The installed memoryguard command exposes these top-level operations:
Command
Purpose
audit [path]
Run a read-only audit and generate a report
open [path]
Open the latest interactive report
explain <finding_id>
Explain evidence and risk for a finding
source <action>
List, add, remove, or preview authorized sources
scan
Scan authorized sources and build the coverage ledger
doctor
Diagnose V2 manifest, domain availability, and native coverage
Install, inspect, pause, repair, or remove host Hooks
provider <action>
Inspect or repair global provider integrations
`storage audit
report`
`storage sweep
compact`
groups <action>
Inspect governed group state
gui [path]
Launch the interactive governance console
desktop
Launch the trusted desktop executor
The old V1 plan, apply, verify, undo, import, and gc workflows may
remain parseable as explicit retired compatibility surfaces, but are not a V1
runtime path. Under V2_ACTIVE they return a stable retired result instead of
writing through a legacy store. Legacy data input is accepted only by the
explicit memoryguard.migration upgrade flow.
Run memoryguard --help or memoryguard <command> --help for the live command
reference.
MCP API
The default MCP discovery surface is intentionally compact. New MCP clients
receive these eleven day-to-day tools through tools/list:
Tool
Purpose
memoryguard_context_bootstrap
Load bounded mandatory rules and relevant memory context
memoryguard_memory_search
Search governed memories by query, lifecycle status, and bounded limit. kind is not an MCP search filter; semantic duplicate/conflict checks are separate advanced governance.
memoryguard_memory_read
Read one governed memory
memoryguard_memory_write
Write and organize a governed memory
memoryguard_memory_update
Update the body, kind, recall policy, or priority of one known memory. It does not change lifecycle status.
memoryguard_memory_delete
Soft-delete a governed memory
memoryguard_memory_status
Inspect shared-memory status
memoryguard_audit
Run a read-only local governance audit
memoryguard_explain
Explain one audit finding and its evidence
memoryguard_capabilities
Discover registered MCP operations and reviewed headless GUI operations with bounded pagination and optional on-demand JSON Schema
memoryguard_invoke
Invoke one discovered MCP or reviewed headless GUI operation; mutating targets require confirmation and an idempotency key
Advanced governance remains available through the GUI and CLI: rule lifecycle,
bindings and shared groups, source scanning, CodeGraph, knowledge and history
review, provider controls, external MCP import, and maintenance operations.
Existing advanced MCP names remain callable for compatibility when an installed
client invokes an exact name, but they are not returned by the default
tools/list. This reduces discovery/schema overhead without removing those
governance capabilities.
memoryguard_capabilities is the discovery path for the broader compatibility
catalog. It supports exact operation lookup, English or Chinese query text,
domain filtering, and offset pagination; schemas are returned only when
include_schema=true is requested for the selected page. The catalog exposes
162 reviewed headless GUI business operations through memoryguard_invoke.
Eight GUI operations remain explicitly restricted by their existing authority:
desktop-only path/folder actions, desktop-admin CodeGraph selection/build, and
SafeBridge protocol actions.
Bounded read responses
MemoryGuard minifies JSON text by default. A replayable read response is capped
at 24,000 UTF-8 bytes across the complete MCP envelope, including every
content block and existing structuredContent. Small responses keep their
existing shape. An oversized read returns a compact receipt with response_ref
and required identifiers; it does not silently truncate the original result.
Fetch a page through the existing broker, after discovering
memoryguard_response_read with memoryguard_capabilities:
Pages are UTF-8 JSON fragments with next_offset; concatenate them in order.
limit is 4–4096 bytes and offsets must be UTF-8 character boundaries. On a
single JSON text payload, fields selects business fields: use a top-level
name or an object-only JSON Pointer such as /data/memory_id. Multi-content
and non-JSON results reject field selection and remain available only as
whole-envelope pages.
Private references live only in the MCP process for at most five minutes: at
most 16 snapshots, each at most 512,000 bytes. They are bound to the exact
trusted session, principal, scope, and active binding revision. Each page
reruns the original read under current authorization and compares its
digest. A denial, changed output, binding/session change, or expired reference
returns a stable refusal such as response_ref_access_denied,
response_ref_expired, or response_ref_result_changed; cached old content
is never used to bypass the current read. Public capability metadata uses its
existing offset pagination. Writes and context bootstrap keep their existing
complete receipt/mandatory-rule contracts and cannot request response
pagination, so a page read never reruns a mutation. If an oversized read cannot
safely create a reference, its bounded receipt reports
delivery.status="unavailable" and action="narrow_query" rather than
promising the whole result can be retrieved.
Example discovery and invocation using the published schemas:
For a mutating target, the invoke envelope must also carry
"confirmed":true and a non-empty "idempotency_key"; the broker forwards
those proofs to the target's existing permission and scope checks.
The underlying compatibility catalog also covers:
governed memory read, search, write, update, delete, and status;
bounded context bootstrap with mandatory-rule isolation;
rule creation, feedback, merge governance, undo, and scope statistics;
Agent binding and shared-group inspection;
source scanning, graph projection, import previews, and build planning;
external MCP discovery and import;
document extraction previews and candidate acceptance;
conversation-history search, timeline, explicit read, export, deletion, and
extraction preview;
provider installation and host-agent enrichment.
Use MCP tools/list for the compact default discovery set. Use
memoryguard_capabilities for the registered compatibility catalog and its
reviewed operation metadata.
Release history: v0.7.9 consolidates canonical governance, local-only token evidence, readable multi-agent governance, and public distribution through GitHub, PyPI, and the official MCP Registry. v0.7.8 records the preceding governance, telemetry, and Codex runtime work; v0.7.7 makes bare provider repair safe in a verified, uniquely bound control home and aligns installed Codex MCP/Hook repairs to the current interpreter while preserving Agent and shared-group identity. v0.7.6 makes Codex Hook/MCP runtime selection consistent through one immutable snapshot, shortens Hook state lock windows, and keeps bootstrap success/failure state honest with explicit mandatory-overflow fail-closed handling. Earlier release records retain the detailed v0.7.5 conflict-review, v0.7.4 canonical-governance, v0.7.3 shared-history, and v0.7.2 write/read and Codex lifecycle changes. The
v0.7.1 V2-only migration and desktop lifecycle work remains documented as
historical release context.
Acceptance boundary: the Graphify evidence is the focused 3 / 3 result
plus the real full-repository export/projection described above. It does not
claim that upstream Graphify's full-repository test suite passed.
Next after release: broader CodeGraph/Skills ingestion, more operator-friendly
maintenance reports, and additional migration observability. Long-term records
are not retired merely because they are old.
Later: team and enterprise capabilities only after validated demand.
Contributing
Issues and pull requests are welcome. Read CONTRIBUTING.md
before submitting a change. Pull requests require agreement to the
CLA.
