dsh-agentmemory

agentmemory for DeepSeek Harness — full memory_* tools, automatic capture hooks, and opt-in context injection over the local REST server.

English · 中文

dsh-agentmemory connects agentmemory, a local memory server for coding agents, to DeepSeek Harness (DSH). It exposes the full memory_* tool set, captures observations automatically from sessions, prompts, and tool activity, and can inject recalled context into the system prompt — all over the local REST server, with no MCP bridge required.

Features

• Full tool surface — all 54 memory_* tools (8 core) mapped to /agentmemory/*, plus memory_observe and a memory_http escape hatch for any endpoint.
• Automatic capture — session, prompt, tool-use, subagent, workflow, and approval activity are recorded as observations in the background.
• Context injection — opt-in; recalled context from the session is added to the system prompt.
• Safe by default — credential redaction, per-call timeouts, non-blocking observation, and destructive tools gated behind a flag.
• No MCP required — no stdio bridge, no child process; the running REST server (localhost:3111) is the only dependency.

Requirements

An agentmemory server running on the same machine (default http://127.0.0.1:3111, viewer on 3113):

curl -fsS http://127.0.0.1:3111/agentmemory/livez
# {"service":"agentmemory","status":"ok"} when healthy

If the server is unavailable, the plugin logs a warning and the memory_* tools return errors; the harness itself keeps running normally.

Install

Native mount

git clone https://github.com/elementor-i/dsh-agentmemory ~/dsh-plugins/dsh-agentmemory

Add an insert patch to ~/.dsh/config.yaml (or $DSH_HOME/cordis.patch.yml for all profiles):

- insert:
    - id: dsh-agentmemory
      name: '$HOME/dsh-plugins/dsh-agentmemory/lib/index.js'

Then restart DSH. Compiled output is committed, so no build step is required.

Official CLI

# npm registry (recommended) — @elementor-i/dsh-agentmemory@^0.1.1
npx -p @deepseek-ai/dsh dsh plugin --profile web add @elementor-i/dsh-agentmemory

# or directly from the source repository
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:elementor-i/dsh-agentmemory

Replace web with your profile name. The package declares a dsh.bundle.patch, so it loads as a profile layer; compiled output is committed, so no build runs on install.

Plugin manager (AI-assisted)

Oh-DSH-Desktop's plugin manager installs with an isolated preview and rollback. It is driven by an AI coding agent through tools — not shell commands — so the fastest path is to paste this prompt to your assistant:

Install dsh-agentmemory from the Oh-DSH-Desktop plugin manager: refresh the plugin catalog, prepare the install as an isolated preview, and apply it after I review the preview.

Under the hood the assistant calls these tools:

desktop_plugin_search  { query: 'dsh-agentmemory', refresh: true }   # confirm it is in the catalog
desktop_plugin_prepare { action: 'install', pluginId: 'dsh-agentmemory' }
desktop_plugin_apply   {}                                            # after you approve the preview

Notes:

• The plugin must already be in the public DSH catalog. The catalog is rebuilt hourly from GitHub repos tagged dsh-plugin, so a newly published repo appears within about an hour.
• desktop_plugin_apply restarts the live DSH runtime — review the isolated preview before applying.

Configuration

All keys are optional and have safe defaults. The environment variables AGENTMEMORY_URL, AGENTMEMORY_SECRET, and AGENTMEMORY_PROJECT_NAME are honored as fallbacks.

dsh-agentmemory:
  baseURL: http://127.0.0.1:3111   # empty -> $AGENTMEMORY_URL -> default
  secret: ""                       # empty -> $AGENTMEMORY_SECRET (Bearer)
  timeoutMs: 10000
  observeTimeoutMs: 3000           # fire-and-forget hook timeout
  registerTools: true
  coreToolsOnly: false             # true -> only the 8 core tools
  dangerousTools: false            # true -> expose destructive/expensive tools
  projectName: ""                  # empty -> $AGENTMEMORY_PROJECT_NAME -> git repo basename
  injectContext: false             # inject recalled context into the system prompt
  injectMaxChars: 4000
  healthCheck: true
  hooks:
    enabled: true
    capturePrompts: true
    captureToolUse: true
    toolNameFilter: []             # non-empty -> only these tool names
    captureSubagents: true
    captureWorkflow: true
    captureApprovals: false
    preCompactSnapshot: false      # approximate PreCompact on request-error
    maxObservationBytes: 8000
    redactSecrets: true

| Key | Default | Description | | --- | --- | --- | | baseURL | http://127.0.0.1:3111 | agentmemory REST base URL | | secret | "" | Bearer secret (falls back to AGENTMEMORY_SECRET) | | timeoutMs | 10000 | per-tool HTTP timeout | | observeTimeoutMs | 3000 | hook observation timeout (fire-and-forget) | | registerTools | true | register the memory_* tool set | | coreToolsOnly | false | register only the 8 core tools | | dangerousTools | false | expose governance_delete, heal, consolidate, reflect, crystallize | | injectContext | false | inject recalled context into the system prompt | | healthCheck | true | warn on startup if the server is unreachable |

Hooks

Activity is captured automatically through DSH's official events. Handlers are non-blocking: requests use a short timeout and are never awaited, and waterfall events always call next().

| agentmemory hook | DSH event | Mode | | --- | --- | --- | | SessionStart | session/created | emit | | UserPromptSubmit | agent/inbox/inserted | emit | | PreToolUse | tools/pre-execute | waterfall | | PostToolUse / PostToolUseFailure | tools/result | emit | | PreCompact (approx) | agent/request-error | waterfall | | SubagentStart / SubagentStop | subagent/start / subagent/end | emit | | Notification | approval/request | waterfall | | TaskCompleted | agent/turn-stopping | serial | | SessionEnd | session/disposed → /session/end | emit | | context injection | system-prompt/assemble | waterfall |

Tools

Core set (registered by default, or exclusively with coreToolsOnly):

memory_save · memory_recall · memory_smart_search · memory_sessions · memory_lesson_save · memory_consolidate · memory_reflect · memory_diagnose

The remaining 46: memory_commits · memory_commit_lookup · memory_compress_file · memory_file_history · memory_timeline · memory_vision_search · memory_lesson_recall · memory_lesson_delete · memory_graph_query · memory_relations · memory_patterns · memory_profile · memory_audit · memory_verify · memory_heal · memory_crystallize · memory_governance_delete · memory_slot_create · memory_slot_get · memory_slot_append · memory_slot_replace · memory_slot_list · memory_slot_delete · memory_action_create · memory_action_update · memory_frontier · memory_next · memory_lease · memory_checkpoint · memory_routine_run · memory_signal_send · memory_signal_read · memory_sentinel_create · memory_sentinel_trigger · memory_sketch_create · memory_sketch_promote · memory_facet_tag · memory_facet_query · memory_mesh_sync · memory_team_share · memory_team_feed · memory_snapshot_create · memory_export · memory_claude_bridge_sync · memory_obsidian_export · memory_insight_list

Extras:

• memory_observe — record a raw observation manually (automatic capture usually covers this).
• memory_http — call any /agentmemory/* endpoint with a JSON body or query (for endpoints without a dedicated tool).

How it works

Three parts cooperate over the local REST server:

• Tools — each memory_* tool maps to the matching /agentmemory/* endpoint.
• Hooks — DSH lifecycle events are forwarded to the server as observations.
• Injection — on session start, the server returns recalled context, which is added to the system prompt when injectContext is enabled.

DSH events     ──▶  /agentmemory/observe        automatic capture
memory_* tools ──▶  /agentmemory/*              on-demand operations
session start  ──▶  context ──▶ system prompt   optional injection

Compatibility

Tool names and endpoints follow agentmemory's official reference. A running server may differ from the latest release:

• some features are disabled by default (for example /slots returns 503 with an enable hint);
• newer endpoints may not be present yet.

On a 4xx or 5xx response, follow the hint in the response body or align the server version. memory_http reaches endpoints without a dedicated tool.

Development

Build from source:

npm install
npm run typecheck
npm run build
npm test        # read-only checks against a running server

FAQ

The plugin manager fails to install — what else can I try?

Oh-DSH-Desktop's plugin manager and the official CLI both end up running dsh plugin --profile <name> add <package>. If the plugin manager fails in your environment, the CLI is an equivalent fallback:

npx -p @deepseek-ai/dsh dsh plugin --profile desktop add @elementor-i/dsh-agentmemory

Replace desktop with your profile name, then restart DSH. If you are managing the desktop profile, run the command with the same DSH_HOME the desktop app uses (on macOS, ~/Library/Application Support/Oh-DSH-Desktop/dsh).

ERR_PNPM_UNEXPECTED_STORE or "pnpm failed in profile directory"

This can happen when the profile's node_modules was linked from a pnpm store that no longer exists — for example, after a plugin-manager preview directory was cleaned up. A workaround observed on macOS is to relink from the current store, then retry:

CI=true dsh plugin --profile desktop install
dsh plugin --profile desktop add @elementor-i/dsh-agentmemory

(CI=true lets pnpm recreate node_modules without an interactive prompt.)

The plugin manager times out on gh

The plugin manager shells out to the GitHub CLI (gh) to resolve commits and clone repositories. If it reports a gh timeout while the same command works in your own terminal, retrying — or using the CLI path above — is usually the quickest way forward.

These notes describe symptoms observed in a specific environment; they are not a guarantee that every setup behaves the same way.

License

MIT © 2026 Element

Acknowledgements

• agentmemory — the memory server this plugin connects to.
• DeepSeek Harness — the host and its event system.