Cordis Memory with DeepSeek Harness Integration

English | 中文

dsh-missher-memory is an independently installable DeepSeek Harness bundle for recovering architecture, decisions, progress, failed approaches, and next steps in long-running projects. The current package includes indexed recall and reversible duplicate consolidation. It neither changes Harness core nor copies or modifies an existing legacy database.

For a fresh computer, use the Cordis service guide for a reusable dsh-missher-memory/core plugin, or the Harness Agent guide for the existing Bundle. Core provides missherMemoryService without Harness services. Brain is optional and only controls Harness automatic recall.

Platform support

The Core is pure JavaScript and imports only Node built-ins at runtime; the Harness adapter uses Host peer packages. CI is configured to build one canonical .tgz and verify the same bytes on macOS Intel, macOS Apple Silicon, Windows x64, and Linux x64. The checks cover Cordis containers, package safety, CLI install/uninstall, and synthetic data. The CLI matrix is pinned to DeepSeek Harness Desktop 0.3.6 / Harness 0.1.1-rc.2. A configured matrix is not evidence that an unpublished candidate passed every platform; consult the current delivery report.

Windows ARM and Linux ARM are not claimed until stable native runners and a shipped Harness target are available. There is no platform-specific database payload or native addon inside the package.

How it works

• memory_search searches external memory and reviewed plugin memory only through the active session's confirmed project binding, with source, time, and a stable reference on each result.
• cwd creates an in-memory binding candidate only. Durable state keeps an irreversible project key, basename, short hash, and encrypted external session identifiers; it never stores an absolute cwd.
• Project memory and personal preferences use separate scopes. Project search cannot read another project, and personal search does not read the external project database.
• Candidate capture defaults on for newly bound projects. Session disposal creates review candidates only after the user explicitly binds the project; candidates never become approved memory automatically.
• Automatic recall defaults on for newly bound projects. It contributes reviewed atoms, reversible capsules, and optional legacy rows to the Desktop Brain Hub; the Hub is the only component that appends one visible, source-attributed recall message.
• Old, unpinned, exact duplicate reviewed atoms are consolidated automatically after seven days. Sources are archived rather than deleted, and rolling back a capsule restores every source and FTS row exactly.
• Missing, damaged, unsafe, or timed-out databases return stable states and fail open without blocking Harness startup or a session.

Install

The plugin requires a DeepSeek Harness 0.1.x Host with Node ^22.19.0 or >=24. The delivered tarball needs no Python, shell script, or native dependency build:

dsh plugin --profile web add https://github.com/Missher12/dsh-missher-memory/releases/download/v0.3.0-cordis.0/dsh-missher-memory-0.3.0-cordis.0.tgz
dsh --profile web --dump-config

The configuration contains both dsh-missher-memory and missher-memory when the bundle patch has been composed into the profile; this does not prove runtime activation. The Harness Bundle uses dshHomePath and tools. The separate Brain adapter waits for missherBrain; manual search remains available without it. Restart a compatible Harness Host and finish the first binding in Settings → Super Memory, then verify that the current Agent can call memory_search.

A fresh Windows or macOS install needs no vectors.db: after explicit project binding, built-in project memory uses the plugin-owned state.db. A vectors.db is only an optional, read-only source for legacy memory. If it is not at $HOME/.local/share/missher-memory/tencentdb/vectors.db, you may set MISSHER_TENCENTDB_DIR to the existing absolute directory that contains it before starting Harness. The plugin never creates a missing directory or empty database and rejects links and escaping paths.

First binding

1. Open a top-level session in the target project so Settings shows its basename and short-hash candidate.
2. Review the source list, which contains record counts and time ranges but no record text, and select only sources that belong to the project.
3. Confirm a new binding or link another worktree candidate to an existing project.
4. Candidate capture and automatic recall start enabled for a newly bound project and remain independently switchable. Existing project settings are never migrated or overwritten.

The legacy database has no trustworthy project id. The plugin never classifies sources from cwd, text similarity, or time. A wrong source choice assigns history to the wrong project, so the first binding requires human review.

Search

The model or user can explicitly call:

memory_search({ query: "packaged smoke", scope: "project", limit: 5 })

scope is project or personal. The query is interpreted as literal text rather than FTS operators, and results are bounded by count and UTF-8 bytes. Search never creates state.db and never triggers candidate capture.

Candidate review and recall

After capture is enabled, the plugin buffers only direct user and assistant text from top-level sessions; it ignores tool output, plugin injections, and delegated sessions. If any message matches a credential, private key, connection string, identity number, financial number, or sensitive user path, the entire session produces no candidate.

Settings lets the user edit, merge, approve, pin, or forget candidates. Only approved memories can be searched or recalled; pinning changes order only. Project deletion removes that project's bindings, settings, candidates, project memory, and personal memory derived from its candidates without touching the external database.

Automatic recall uses only reviewed content and explicitly bound external sources. It has an independent switch, result limit, and byte budget. Errors, timeout, or invalid state contribute nothing. The optional legacy reader code is bundled, but no legacy database, user state, credential, or path is included in the package.

Data and uninstall

Plugin-owned state lives under $DSH_HOME/missher-memory/, primarily in permission-restricted state.db and a local key. Candidate and approved text remain readable in state.db; project aliases are irreversible digests and external session identifiers are encrypted with the local key. DATA-RETENTION.md defines the complete retention rules and SECURITY.md defines the threat model.

Export or delete projects in Settings as needed, then uninstall:

dsh plugin --profile web remove dsh-missher-memory
dsh --profile web --dump-config

Uninstall removes the bundle and profile patch but preserves $DSH_HOME/missher-memory/ by default for reinstall recovery. After confirming backups and retention requirements, the user may delete that exact directory. Do not delete or move the external vectors.db.

Status reference

• not connected (optional): the legacy directory or vectors.db is absent; built-in project memory remains available and the plugin does not create an external database.
• unsafe path: the directory, database, or plugin state is a link, a non-regular file, or fails containment checks.
• incompatible: the external tables, FTS5 schema, or plugin state schema is unsupported.
• corrupt: SQLite cannot validate the database read-only.
• timeout: the Worker was terminated and will be rebuilt for the next search.

Before distribution, run:

node scripts/verify-package.mjs dist/dsh-missher-memory-0.3.0-cordis.0.tgz
node scripts/native-smoke.mjs --archive dist/dsh-missher-memory-0.3.0-cordis.0.tgz

native-smoke.mjs uses synthetic data only. Passing --cli /absolute/path/to/dsh-cli.js additionally installs, composes, and removes the tarball in a temporary profile.

Cordis prerelease status

0.3.0-cordis.0 is a prerelease. Download the prebuilt package and check version-specific verification in the release notes. Store inclusion is separate and depends on the curated registry accepting the entry.

The Cordis upgrade removes the mandatory Brain dependency. The earlier 117-test maintenance evidence below predates this upgrade; see CORDIS.md and the current delivery report for fresh verification. CLI install/remove evidence is separate from runtime activation. The packaged smoke reports runtimeMode: cordis-with-synthetic-host-services and realHostActivationVerified: false, and checks reinstall restoration using temporary synthetic data. No real Desktop Brain or UI acceptance is claimed by this smoke.

Reviewed memory remains untrusted historical data, never new authorization. Forgetting removes derived atoms and capsules, while retaining the forgotten candidate for review history. Project JSON export currently excludes capsules and archived atoms and is not a full backup.